trw-memory is a persistent memory engine for AI agents: an agent memory layer that gives LLM agents long-term memory across sessions, stored locally in SQLite. Use it as an async Python SDK, a CLI, or an MCP memory server. The core install recalls with keyword search; optional extras add hybrid retrieval (BM25 + dense vectors via sqlite-vec, fused with Reciprocal Rank Fusion) and cross-encoder reranking, alongside lifecycle scoring, tiered storage, and a knowledge graph. It is the standalone memory backend of TRW Framework and works without it.
Release status: Alpha and source-available under BSL 1.1. The public API may change before 1.0; evaluate upgrades in a test environment before production rollout.
Why trw-memory · Quick start · Python API · Conversation memory · CLI · Benchmarks · mem0 comparison · MCP server · Security and network behavior · FAQ · Development
TRW-Memory is a standalone persistent memory engine for AI agents that gives coding agents searchable, long-lived knowledge storage. It stores learnings (patterns, gotchas, architecture decisions) in SQLite with optional YAML backup, and retrieves them using hybrid search that combines keyword matching (BM25) with dense vector similarity. It also stores conversation memory: store_conversation() keeps chat turns verbatim so they can be recalled later, RAG-style, as evidence for an answer.
Designed as the storage backend for trw-mcp and TRW Framework, but usable independently by any AI agent framework that needs persistent memory with recall.
- Local-first. With the default configuration all data lives in a local SQLite store (plus an optional YAML sidecar). There is no usage tracking or content phone-home; the only network-capable surfaces are optional model downloads and opt-in remote sync. See Telemetry and network behavior.
- No generative LLM call at write time.
store_conversation()stores every turn verbatim with its date and the turn it replied to. Ingest never calls a generative model (the optional local embedding model still encodes each turn); recall does the work. - Hybrid retrieval, optional. With the retrieval extras installed: BM25 (with stemming) + dense vectors via sqlite-vec + Reciprocal Rank Fusion + a cross-encoder re-ranker. Without them, retrieval degrades gracefully to the backend's built-in keyword search.
- Measured on public benchmarks. With no LLM in the loop, the gold evidence turn lands in the top 10 for 85.7% of LOCOMO questions (n = 1,540) and 93.8% of LongMemEval_S questions (n = 470); top 50: 92.7% and 97.4%. Write and recall costs stay flat as the store and the number of projects sharing it grow. See Benchmarks for method and caveats.
- Works offline.
TRW_OFFLINE=1/HF_HUB_OFFLINE=1block model downloads;local_only: truehard-blocks all remote sync and model download. Hybrid retrieval offline needs the models already in the local cache. - MCP memory server included.
trw-memory-serverexposes store, recall, search, consolidate, forget and more as MCP tools over stdio, or over a per-user loopback HTTP daemon. - Evaluating a mem0 alternative? Using mem0's open-source evaluation suite on one LOCOMO conversation (n = 152 questions per system, one run each, local
llama3.1as answerer, judge and mem0's extraction model), neither paired test detected a statistically significant accuracy difference between trw-memory and mem0 (OSS), and trw-memory called no generative LLM to ingest the conversation. That does not establish equivalence or superiority; read the numbers and caveats first. - Source-available. BSL 1.1, alpha; the public API may change before 1.0.
- MemoryClient SDK -- High-level async Python client with store/bulk_store/store_many/recall/search/search_fts/forget plus audit_learning and review_quarantined
- Hybrid Search (BM25 + vector + re-ranker) -- BM25 keyword matching + dense vector similarity via sqlite-vec, combined with Reciprocal Rank Fusion (RRF), then a cross-encoder re-rank with a confidence floor that scales with the requested limit. Learn more
- FTS5 keyword search --
MemoryClient.search_fts()runs indexed SQLite FTS5 keyword search with BM25 ranking over content/detail/tags for pure-keyword queries that don't need hybrid ranking; degrades to an empty result when FTS5 is unavailable - Hybrid order preservation by default -- recall preserves the hybrid BM25+dense+RRF order when enough local candidates are already available, avoiding a legacy score-scale mismatch in tier merging. To restore the legacy tier rescore for a workload, set
MEMORY_RECALL_PRESERVE_HYBRID_ORDER=false. - Tiered Storage -- Hot/warm/cold tiers for fast recall, warm-sidecar persistence, recall-time cold promotion, and explicit sweep-based archiving/purging. Architecture details
- Semantic Deduplication -- Detects and merges near-duplicate learnings using cosine similarity (0.85 threshold)
- Knowledge Graph for AI -- Tag co-occurrence and similarity edges, BFS traversal, importance boost/decay, cross-validation propagation. Docs
- Memory Consolidation -- Episodic-to-semantic consolidation via clustering with the current shipped path using heuristic/fallback summarization
- Outcome-based memory scoring -- outcome-driven utility (Q-value) scoring with EMA updates, Ebbinghaus forgetting curve applied at query time, Bayesian MACLA calibration
- Remote Sync -- Publish/fetch learnings across installations with vector clock conflict resolution and SSE live updates
- Security -- optional AES-256-GCM field encryption (off by default), PII detection with publish-time masking, memory-poisoning anomaly detection (z-score; enforcement is opt-in), RBAC, audit trail. See Security defaults
- Agent Integration --
register_tools()for agents that expose aregister_tool()ortool()API,@auto_recalldecorator - Framework Integrations -- VS Code interface contract and an OpenAI-compatible adapter
- CLI -- Full command-line interface for store, recall, search, forget, consolidate, export/import
- MCP Tools -- store, recall, search, consolidate, forget, status, audit, review, wiki-lint, and an explicit code index (index/search/symbol) — served by
trw-memory-server - Dual Storage Backends -- SQLite with keyword search (primary) + YAML (backup) with one-time migration
trw-memory is the standalone memory engine for TRW (The Real Work) — a methodology layer for AI-assisted development that provides stateless agents with a persistent memory layer designed to enable self-improvement across sessions via knowledge compounding. Cross-session recall is measured to let an agent finish work it otherwise cannot; broad coding-task lift is not established, and the measured SWE-bench Verified result is unfavourable (56 vs 79 of 112 paired-valid problems). See the verification docs for the current methodology and evidence posture. It works alongside trw-mcp, the MCP server that builds its tooling on this engine.
- trw-memory (this repo): Standalone AI agent memory engine with hybrid retrieval, scoring, and lifecycle
- trw-mcp: MCP server for AI coding agents — uses trw-memory as its backend
# Core local engine (SQLite + built-in keyword search)
pip install trw-memory
# Recommended hybrid retrieval
pip install "trw-memory[embeddings,vectors,bm25]"
# The full retrieval stack (same as the line above, one name)
pip install "trw-memory[all]"By default, memories are stored in .memory/ relative to the current directory. Override with MEMORY_STORAGE_PATH env var.
For source development, clone the repository and run pip install -e ".[dev]" from trw-memory/. Tested on CPython 3.10 through 3.14; see Platform and interpreter notes for SQLite engine details.
import asyncio
from trw_memory.client import MemoryClient
async def main() -> None:
async with MemoryClient(namespace="project:my-app") as client:
await client.store(
"Pydantic v2 requires use_enum_values=True for YAML round-trip",
tags=["pydantic", "gotcha"],
importance=0.8,
)
# Uses hybrid retrieval when the optional rankers are installed.
results = await client.recall("pydantic serialization", limit=10)
high_impact = await client.search(min_importance=0.7, tags=["gotcha"])
print(results, high_impact)
asyncio.run(main())MemoryClient also provides store_many() and bulk_store() for batch writes, search_fts() for keyword-only lookup, forget() for deletion, and audit_learning() / review_quarantined() for lifecycle and security workflows.
# inside `async with MemoryClient(...) as client:`
turns = [
{"role": "user", "speaker": "Caroline", "content": "I went to a LGBTQ support group yesterday."},
{"role": "assistant", "speaker": "Melanie", "content": "That's great! What did it look like?"},
]
summary = await client.store_conversation(turns, observed_at="2023-05-08T13:56:00+00:00", session_id="s1")
rows = await client.recall("what did the support group look like", limit=5)store_conversation() stores every turn verbatim and carries the preceding
context_turns (default 1) of the same conversation alongside it, so a reply
like "What did it look like?" is retrievable by what it was replying to. No
generative LLM is called at ingest time (the optional local embedding model
still encodes each turn): the raw turn, its date and its neighbourhood are
the evidence, and the reader does the inference at recall time. Feeding a
conversation in chunks? Pass the last turns you already stored as
preceding=.
from trw_memory.client import MemoryClient
client = MemoryClient(namespace="project:my-app")
# Register tools with any agent that has register_tool() or tool() API
client.register_tools(agent)
# Or use the auto_recall decorator
@client.auto_recall(query_from="prompt")
async def handle_prompt(prompt: str, recalled_memories: list | None = None) -> str:
# recalled_memories is automatically injected with relevant context
recalled_memories = recalled_memories or []
return f"Found {len(recalled_memories)} relevant memories"# Store a learning
trw-memory store "Always use connection pooling for PostgreSQL" --tags db,performance --importance 0.8
# Recall by query
trw-memory recall "database optimization" --limit 5
# Search with filters
trw-memory search --tags security --min-importance 0.7
# Consolidate related entries
trw-memory consolidate --namespace project:my-app --dry-run
# Export/import a namespace's entry data
trw-memory export --format json > memories.json
trw-memory import memories.json --namespace project:new-app
# Forget an entry by ID
trw-memory forget M-abc12345 --namespace project:my-app
# Re-encode stored vectors after an embedding-model change (idempotent, resumable)
trw-memory reembed --namespace project:my-app
# Rebuild the SQLite DB from the cold YAML tier or a snapshot
trw-memory restore --from-cold
trw-memory restore --from-snapshot latest
# Snapshot management (VACUUM INTO rotation)
trw-memory snapshot create --tier daily
trw-memory snapshot list
trw-memory snapshot rotate
# Lint wiki page JSON for missing targets/backlinks/provenance
trw-memory wiki-lint pages.json
# Explicit code index: index, lexical search, and symbol lookup
trw-memory code-index ./src
trw-memory code-search ./src "hybrid_search" --language python --limit 5
trw-memory code-symbol ./src MemoryClient
# Status overview
trw-memory statusExport enumerates the requested namespace (default unless specified), not the
whole project or every namespace. Use an unchanged store for a consistent export:
pagination is not a snapshot across concurrent writes. JSON/YAML output retains
the entry format and is materialized in memory; it is not a streaming database
backup and does not include arbitrary project files or stored vector indexes.
The separate snapshot commands above serve database snapshot management.
from trw_memory.storage.sqlite_backend import SQLiteBackend
from trw_memory.models.memory import MemoryEntry
backend = SQLiteBackend(db_path=".trw/memory.db")
entry = MemoryEntry(id="M-abc12345", content="Use WAL mode for concurrent readers", namespace="default")
backend.store(entry)
results = backend.search("query", top_k=10, namespace="default")Every number below is reported with its sample size; confidence intervals and paired tests are given where they were computed. Apart from the mem0 comparison, these are same-harness ablations — retrieval strategies compared on one fixed corpus and query set — not leaderboard claims against other systems. The framework's evidence posture is described in the verification docs; the mem0 comparison's method and scripts live in benchmarks/locomo/.
Scope first. One LOCOMO conversation of ten, one run per system. The answerer, the judge and mem0's extraction model were all a local 8B
llama3.1; that judge was not calibrated against the GPT-class judges behind mem0's published numbers, so compare the two columns with each other, not with mem0's website. mem0 was run as its open-source SDK (mem0ai2.0.20), not Mem0 Cloud. Results apply to these configurations only.
We ran mem0's open-source evaluation suite (commit 4b61c5d) unmodified against both systems: same dataset parsing, same answer prompt, same LLM judge, same cutoffs, and the same embedding model (all-MiniLM-L6-v2) for both. Conversation 0, n = 152 questions per system, paired by question, trw-memory 0.19 defaults:
| Memories given to the answerer | mem0 (OSS) | trw-memory | McNemar p |
|---|---|---|---|
| top 10 | 88.2% [82.1, 92.4] | 91.4% [85.9, 94.9] | 0.38 |
| top 50 | 92.1% [86.7, 95.4] | 91.4% [85.9, 94.9] | 1.00 |
Neither paired test detected a statistically significant accuracy difference. That does not establish equivalence or superiority; it means this sample could not tell the two apart.
Ingestion measurements for the same run (419 turns, one machine, single run):
| mem0 (OSS) | trw-memory | |
|---|---|---|
| Generative LLM calls during ingestion | ~2 per turn | none |
| Ingestion wall-clock time | 1 h 28 min | ~75 s |
| Storage approach | LLM-extracted facts | the supplied turns, verbatim, with their dates and the turn they replied to |
These measurements do not establish total operating cost: recall and answer generation are not included, and "top k" counts stored items, not equal token budgets (a verbatim turn and an extracted fact are different units). Verbatim storage avoids generative rewriting during ingestion; it does not guarantee correct input metadata, retrieval, or answers.
How trw-memory gets there without a generative model at write time: store_conversation() keeps each turn verbatim with its conversational context, and recall does the work (BM25 with stemming + dense vectors + rank fusion + a cross-encoder re-ranker that drops low-confidence rows). Evidence retrieval over all ten LOCOMO conversations (n = 1,540 questions, no LLM in the loop): the gold evidence turn is in the top 10 for 85.7% of questions and in the top 50 for 92.7% (mean reciprocal rank 60.3).
A second, independent LLM-free benchmark runs the same way: LongMemEval_S (cleaned, HF revision 98d7416c), n = 470 non-abstention questions, each with its own haystack of 38-62 sessions. An answer-bearing turn is in the top 10 for 93.8% of questions (Wilson 95% CI 90.6-95.1, first run of this configuration) and in the top 50 for 97.4%; the answer session is in the top 10 for 93.3%. Weakest question type: single-session preference, 73.3% (n = 30). Harness: benchmarks/longmemeval/.
Both benchmarks are the inner loop for retrieval changes, and a change ships only if it is non-inferior on both. Measured that way, paired question-by-question against the previous defaults:
| Change | LOCOMO (n = 1,540) | LongMemEval (n = 470) |
|---|---|---|
bge-small-en-v1.5 replaces all-MiniLM-L6-v2; session dates indexed with each turn |
hit@10 84.0% -> 85.1% (McNemar p = 0.044) | not run for this change |
| Entity-bridge second hop after re-ranking | hit@10 85.1% -> 85.6% (p = 0.022); multi-hop recall@10 49.9% -> 51.3% (p = 0.020) | no question changed outcome |
| Confidence floor scales with the requested limit | hit@50 92.5% -> 92.7% (p = 0.25, not significant) | hit@50 95.3% -> 97.4% (p = 0.002); rows returned at limit 50: min 5 -> 25 |
Measured and rejected on the same harnesses, so they are not in the product: CombMax fusion, larger candidate pools, a different RRF constant, and MMR / Dartboard diversity re-ranking (no significant gain on both; diversity cut LOCOMO temporal recall), plus pseudo-relevance feedback, larger cross-encoders, bge-base and mxbai-embed-xsmall.
Published LOCOMO numbers for mem0, Zep, Memobase, MemOS and LightMem are LLM-judged answer accuracy for a whole retrieve-then-generate pipeline. The retrieval rates above measure something narrower: whether the gold evidence reached the top k, with no model in the loop. They are not comparable, and a table placing them side by side would mislead. Two further cautions: the LOCOMO judged-accuracy literature is actively disputed between vendors, and in mem0's own paper a full-context baseline with no memory at all (72.9%) scores above mem0 itself (66.9%). The only like-for-like comparison here is the paired table above, where both systems ran inside mem0's harness with the same answerer and judge.
On a gold set of real engineering learnings (n = 889 typed queries), Reciprocal Rank Fusion of BM25 + dense vectors outranks either single ranker (the table below is this gold set; point estimates, no intervals computed):
| Retriever | Recall@10 | nDCG@10 |
|---|---|---|
| BM25 only | 0.869 | 0.771 |
| Vector only | 0.914 | 0.806 |
| Hybrid (BM25 + vector, RRF) | 0.938 | 0.839 |
The same direction was observed on a second, independent benchmark (LongMemEval_S, n = 500 questions); those figures are not reproduced here. Fusion earns its keep on the hard questions: exact-match queries are near ceiling for every retriever, so the lift concentrates in the temporal / multi-session discrimination band.
On TRW's own active learning store (n = 175 near-duplicate "rediscoveries"), the share of duplicates a recall would have caught before re-deriving them — the Preventable Rediscovery Ratio — is far higher for hybrid than for keyword search alone, with non-overlapping 95% CIs:
| Retriever | Preventable Rediscovery Ratio (95% CI) |
|---|---|
| BM25 only | 0.720 [0.649, 0.781] |
| Hybrid | 0.943 [0.898, 0.969] |
Hybrid retrieval surfaced more of these previously stored duplicates; this evaluation did not measure whether agents then avoided re-deriving them.
On a controlled recall-dependent benchmark (H1-MEMORY-BENCH), agents with memory solved every task that required recalling a fact established in an earlier session — 58/58 — while agents without memory solved 0/50 (the fact is absent by construction). Paired McNemar p = 3.6×10⁻¹⁵ across 49 matched pairs (exceeds the pre-registered n ≥ 30), replicated on a second model family.
Scope, honestly. This demonstrates the mechanism: cross-session recall lets an agent complete work it otherwise cannot. Whether that compounds into broad, end-to-end coding-task improvement is a separate question, and the measured answer so far is unfavourable: on SWE-bench Verified, TRW solved 56 of 112 paired-valid problems against the baseline's 79 (McNemar p = 6.6×10⁻⁵). That surface is treated as contaminated and settles nothing in either direction, but no general outcome-lift claim is supported. See the verification docs for the full evidence posture.
Recall and write paths were profiled and the terms that scaled with store size were removed. One Apple-silicon machine, single runs under load, before and after the 2026-09 changes:
| Operation | Before | After |
|---|---|---|
| Ingest per row, 4th LOCOMO conversation into a shared store | 1,464 ms | 15 ms |
| Store one entry with 20 sibling project namespaces present | 708 ms | 15 ms |
| Access-time sidecar write per recall, 5,000 / 20,000 rows | 144 / 289 ms | 1.2 / 1.3 ms |
| Warm-row scan per recall, 5,000 / 20,000 rows | 11.3 / 60.5 ms | 0.3 / 3.2 ms |
| Store + background graph enrichment, rows 901-1,200 | 15.85 +/- 0.98 ms | 8.38 +/- 0.79 ms |
| Similarity-edge enrichment at 10,000 rows | 12.4 ms | 2.3 ms |
The shape matters more than any single row: these costs used to rise with the number of stored rows and with the number of projects sharing a store, and are now flat. Similarity edges are searched over the whole namespace instead of the newest 500 rows, and bytes written per recall fell from megabytes to about 18 KB.
Throughput (historical single-run baseline, not CI-backed, measured before cross-encoder re-ranking became the default; re-ranking adds roughly 30-300 ms per recall on CPU): sub-millisecond store (p95 ≈ 0.31 ms) and ~116 ms hybrid recall p95 at 1,000 entries; on-disk footprint ≈ 1.2 MB per 1k entries.
The engine is organized as a set of focused subpackages under src/trw_memory/. (For the
authoritative, always-current layout, browse the source tree directly — file-level listings
drift quickly.)
| Path | Responsibility |
|---|---|
client.py (+ _client_*.py) |
MemoryClient SDK — the recommended entry point; store/recall/search/forget/bulk + lifecycle/tiering/org-shared helpers |
cli.py, cli_parser.py, cli_*.py |
trw-memory command-line interface and its formatters/storage helpers |
server.py, tools/ |
FastMCP server entry point and the MCP tool implementations (fastmcp is a core dependency) |
storage/ |
SQLite primary backend (WAL, sqlite-vec vectors, snapshots, recovery, resilient fetch) + YAML backend, behind a shared StorageBackend interface; _dbapi.py driver shim |
retrieval/ |
BM25 sparse, dense vector, RRF fusion, and the hybrid_search() pipeline + admission/source policies and token budgeting |
lifecycle/ |
Utility scoring (Q-learning, Ebbinghaus decay, Bayesian calibration), semantic dedup, consolidation, anchor validation, and tiers/ hot/warm/cold management |
graph.py (+ _graph_*.py) |
Knowledge graph — similarity/tag edges, BFS traversal, clusters, conflicts, cross-project, decay |
bandit/ |
Bandit selectors (Thompson, contextual, change-detection) for adaptive ranking |
code_index/, wiki/ |
Explicit code index (chunker/indexer/symbols/search) and wiki page indexing + lint |
embeddings/ |
Embedding provider protocol + local sentence-transformers provider |
sync/ |
Remote publish/fetch with vector clocks, three-way merge, retry queue, SSE subscriber |
security/ |
AES-256-GCM field encryption, PII detection/redaction, poisoning/anomaly defense, RBAC, provenance, audit, trust scoring, quarantine |
integrations/, adapters/ |
VS Code integration (plus the adapter factory) and an OpenAI-compatible adapter |
models/, namespaces/, migration/, utils/ |
Pydantic models/config, namespace lifecycle + validation + path mapping, YAML→SQLite migration, and shared utilities |
| Name | Module | Description |
|---|---|---|
MemoryClient |
client |
High-level async SDK — store, bulk_store, store_many, recall, search, search_fts, forget, audit_learning, review_quarantined, register_tools, auto_recall |
SQLiteBackend |
storage.sqlite_backend |
Primary storage with keyword search, WAL, and sqlite-vec vectors |
YAMLBackend |
storage.yaml_backend |
File-based storage (backup/migration) |
hybrid_search() |
retrieval.pipeline |
BM25 + dense vector search with RRF fusion |
bm25_search() |
retrieval.bm25 |
BM25Okapi sparse keyword retrieval |
dense_search() |
retrieval.dense |
Cosine similarity vector search |
rrf_fuse() |
retrieval.fusion |
Reciprocal Rank Fusion combiner |
KnowledgeGraph functions |
graph |
Tag/similarity edges, BFS traversal, decay |
TierSweepResult |
lifecycle.tiers |
Hot/warm/cold sweep, promote, demote, purge |
DedupResult |
lifecycle.dedup |
Duplicate detection (skip/merge/store decisions) |
compute_utility_score() |
lifecycle.scoring |
Q-learning + Ebbinghaus + Bayesian scoring |
MemoryConfig |
models.config |
Configuration via env vars or dict |
MemoryEntry |
models.memory |
Core data model for stored memories |
SQLite (recommended) -- Fast, transactional, supports keyword search, knowledge graph edges, and optional sqlite-vec vector similarity:
from trw_memory.storage.sqlite_backend import SQLiteBackend
backend = SQLiteBackend(db_path=".trw/memory.db")
# Supports: store, get, update, delete, search, count, list_entries,
# list_namespaces, upsert_vector, search_vectorsYAML -- Human-readable, git-friendly, used as backup during migration:
from trw_memory.storage.yaml_backend import YAMLBackend
backend = YAMLBackend(entries_dir=".trw/learnings")The hybrid search pipeline combines sparse keyword retrieval with dense semantic search — ensuring strong results for both exact-match queries and conceptually similar queries. Read the full architecture docs.
Query --> BM25 (keyword, rank-bm25) --+
+--> RRF Fusion (k, configurable) --> Ranked Results
Query --> Dense (cosine, sqlite-vec) --+
BM25 drops function words from the query and suffix-stems tokens on both sides ("researched" meets "research"); after fusion a cross-encoder re-ranks the top recall_rerank_candidates on every MemoryClient.recall() and drops rows it scores below -8, except the top max(5, ceil(limit / 2)), which are always kept (adaptive_rerank_floor: 5 rows at the default limit=10, 25 at limit=50). There is no switch to turn re-ranking off; it is skipped only when sentence-transformers or the cached model is unavailable, in which case recall keeps fusion order. The RRF constant k is configurable via MemoryConfig.rrf_k (env MEMORY_RRF_K); the shipped default is tuned by the memory meta-harness loop and may change between releases, so treat the exact value as a default rather than a contract.
The pipeline gracefully degrades: if BM25 is unavailable, only dense search runs (and vice versa). If neither is available, falls back to the storage backend's built-in keyword search (case-insensitive LIKE matching).
Learning utility is computed from multiple signals. Full scoring documentation:
- Q-learning: Exponential moving average updated from outcome events (success/failure/mixed)
- Ebbinghaus forgetting curve: Time-based Ebbinghaus decay applied at query time (not mutated in storage) — entries naturally fade unless reinforced by recall
- Access recency boost: Recently accessed entries score higher
- Impact score: Author-assigned importance (0.0-1.0)
- Bayesian calibration: MACLA calibration for impact score accuracy
Hot/warm/cold tiering keeps frequently-used memories fast and archives stale ones. Architecture overview:
| Tier | Criteria | Storage | Latency |
|---|---|---|---|
| Hot | Recently recalled entries | In-memory LRU cache | <1ms |
| Warm | Active entries mirrored into the tier runtime | SQLite + JSONL sidecar with full entry payloads | <50ms |
| Cold | Archived entries matched by recall or explicit sweep policy | YAML archive (partitioned by year/month) | <200ms |
The latency column is the design target for the tier lookup itself, not end-to-end recall latency (hybrid recall with re-ranking is slower; see Benchmarks). Store/recall operations keep Hot/Warm in sync, Cold-tier hits are promoted back to Warm within the same recall, and TierManager.sweep() applies the configurable archive/purge policy when callers trigger a lifecycle sweep.
| Feature | Implementation |
|---|---|
| Field encryption | AES-256-GCM with HKDF-SHA256 per-namespace key derivation |
| PII detection | Regex patterns (email, phone, SSN, credit card, API keys) + Shannon entropy analysis. Store path blocks API-key/token writes and records every other detection as metadata — it does not rewrite your stored text. Masking happens at the publish boundary (strip_pii), where the local copy still holds the original |
| Poisoning defense | Z-score anomaly detection on frequency, size, and content patterns — observe mode by default (records + telemetry, does not quarantine); enforce is opt-in |
| Access control | Role-based (admin/editor/viewer) per namespace |
| Audit trail | Append-only security event log |
| Key management | Master key derivation, per-namespace keys, rotation support |
The MCP server ships with the core install (fastmcp is a core dependency):
trw-memory-server # Starts MCP server (stdio transport)To wire it into an MCP client (Claude Code, Cursor, Claude Desktop and others use this shape):
{
"mcpServers": {
"memory": { "command": "trw-memory-server" }
}
}| Tool | Purpose |
|---|---|
memory_store |
Store entry with optional embedding/vector persistence |
memory_recall |
Hybrid retrieval with optional graph traversal |
memory_search |
Filter-based listing (tags, importance, date range) |
memory_forget |
Delete entries by ID or bulk search query |
memory_consolidate |
Trigger episodic-to-semantic consolidation |
memory_status |
Backend stats, entry counts, tier distribution |
memory_audit |
Provenance + lifecycle audit data for one entry |
memory_review |
Approve/reject a quarantined entry |
memory_wiki_lint |
Lint wiki pages for missing targets, backlinks, provenance gaps |
memory_code_index |
Index source code into the explicit code index |
memory_code_search |
Lexical search over indexed code chunks |
memory_code_symbol |
Look up symbols in the explicit code index |
trw-memory-server serve http runs one process per operating-system user, serving
the same MCP tool surface over streamable-http on 127.0.0.1 with a per-user bearer
token. The port is ephemeral by default and published in a 0600 daemon.json beside
the store, so clients discover it rather than hardcode it.
Trust boundary: one principal. The daemon authenticates the token file, not the
caller. Anyone who can read ~/.trw/memory/daemon-token is fully authorized for
every namespace in that store; a namespace argument selects scope, not permission.
The boundary is therefore the user account, and that is deliberate — this transport
is for one user's agents and applications, not for mutually distrusting tenants.
Concurrency: four workers. Each served memory_recall, memory_store and
memory_maintain call runs its synchronous work in a bounded thread pool
(OFFLOAD_MAX_WORKERS = 4 in daemon/_offload.py), opening and closing its own
SQLite connection inside the worker. Four calls make progress at once; the fifth
queues, and that queue is unbounded. A request that is cancelled after it starts
still runs to completion — the result is discarded, not the work.
Shutdown. SIGTERM and SIGINT drain the worker pool, remove the discovery record, and then let the signal take its default disposition, so a service manager stopping the daemon does not leave clients pointed at a dead endpoint. The record is only ever removed when it names this process and the start time this process wrote, so a slow exit cannot delete a successor's record.
Maintenance. A daemon has no session end, so decay, consolidation and WAL
checkpointing never run on their own. memory_maintain(namespace) triggers them and
records last_attempted_at / last_maintained_at per namespace in maintenance.json
beside the store. Scope is not uniform: consolidation is namespace-scoped, while the
decay pass and the WAL checkpoint act on the whole store.
Recall is bounded. Each namespace contributes at most
max(limit * 5, hybrid_search_candidate_pool_size) entries (default 1000) to a
search, chosen as the most recently updated rows. On a larger namespace, older
entries are not searched, and an empty result is not evidence of absence. Raising
MEMORY_HYBRID_SEARCH_CANDIDATE_POOL_SIZE widens it at a real cost: measured on a
6500-row namespace, warm recall was 139.6 ms at 1000 and 1045.8 ms at 10000.
trw-mcp is the MCP server layer of TRW Framework — it exposes a suite of tools, skills, and agents to Claude Code and other AI coding tools (see the trw-mcp README for current counts). trw-memory serves as its memory backend:
trw_learndelegates toSQLiteBackend.store()viamemory_adapter.py(YAML dual-write as backup)trw_recalldelegates toSQLiteBackend.search()/list_entries()as the sole query path- Scoring functions (
compute_utility_score,update_q_value,apply_time_decay,bayesian_calibrate) are canonical in trw-memory and re-exported by trw-mcp - One-time YAML-to-SQLite migration runs automatically on first access
- Optional vector search via
LocalEmbeddingProvider+rrf_fusewhensentence-transformersis installed
Read more about the full TRW Framework architecture.
trw-memory is local-first: with the default configuration all data lives in a local SQLite store (and an optional YAML sidecar). It makes no outbound network calls except the optional model downloads below (embedding model and cross-encoder re-ranker). There is no usage tracking or content phone-home.
| Surface | When | Default | Opt-out / control |
|---|---|---|---|
| Embedding model download | Only when the embedding model — BAAI/bge-small-en-v1.5 by default (33M parameters, 384-dim, about 130 MB of weights; set MEMORY_EMBEDDING_MODEL to change it) — is not already complete in your local Hugging Face cache. A complete cached snapshot makes zero huggingface.co requests — the loader probes the cache before deciding, and forces local_files_only=True unconditionally when the snapshot is complete (only with the [embeddings] extra installed) |
enabled when the extra is present | TRW_OFFLINE=1 / HF_HUB_OFFLINE=1, or local_only: true (alias memory_local_only) — forces local_files_only so no download is attempted; a disclosure log line precedes any network-capable load |
| Cross-encoder model download (re-ranker, on by default since 0.19.0) | Only when cross-encoder/ms-marco-MiniLM-L-6-v2 is not in your local Hugging Face cache and the [embeddings] extra is installed; the same offline switches force local_files_only=True, in which case an uncached model means recall keeps fusion order (no download, no error) |
enabled when the extra is present | TRW_OFFLINE=1 / HF_HUB_OFFLINE=1, or local_only: true; a disclosure log line precedes any network-capable load |
| Remote sync / publish | Only when sync_enabled=true AND local_only=false |
off (sync_enabled defaults false) |
leave sync disabled, or set local_only: true to hard-block all egress |
A warm cache performs no Hub request, and embedding egress is independent of the consent flags. A fetch is attempted only when the cached snapshot is incomplete or absent and no offline switch is engaged; in exactly that case one structured disclosure log names the host and the switch that would block it. learning_sharing_enabled and platform_telemetry_enabled govern learning-content publishing and usage telemetry respectively — neither gates the embedding model fetch. Embedding egress is governed by the local cache, the offline switches, and local_only.
sync_enabled defaults false, so the engine performs no remote sync out of the box even though local_only itself defaults false. Setting local_only: true is the hard-block: an @model_validator forces sync_enabled=False, clears sync_namespace/platform_url, and pins rbac_mode="local", so no remote-capable surface can be re-enabled while it is set.
With an offline switch engaged (TRW_OFFLINE / HF_HUB_OFFLINE) or local_only: true, the standalone engine loads the embedding model with local_files_only=True; if the model is not already cached it raises a clear LocalOnlyViolationError telling you how to pre-download (this is the behaviour in both cases — local.py does not silently fall back to keyword-only recall here). The graceful "degrade to keyword-only, no crash" path is provided one layer up by trw-mcp's embedder wrapper, which catches that error; a direct trw-memory caller that wants keyword-only recall under an offline switch should pre-download the model or run without the [embeddings] extra installed.
| Variable | Purpose | Default |
|---|---|---|
TRW_OFFLINE |
Master offline switch — blocks the huggingface.co embedding-model and re-ranker model downloads | unset |
HF_HUB_OFFLINE |
Upstream huggingface_hub offline switch — also honored | unset |
MEMORY_EMBEDDING_MODEL |
Sentence-transformers model used for dense vectors. Changing it leaves stored vectors in the old model's space until trw-memory reembed re-encodes them (see Upgrading from all-MiniLM-L6-v2) |
BAAI/bge-small-en-v1.5 (33M params, 384-dim, ~130 MB) |
MEMORY_* |
Engine knobs validated by MemoryConfig (e.g. MEMORY_LOCAL_ONLY, MEMORY_EMBEDDING_TRUST_REMOTE_CODE, retrieval + lifecycle tuning) |
per-field |
| Capability | Default | Notes |
|---|---|---|
| Field-level encryption | off (encryption_enabled=False) |
opt-in (AES-256-GCM per-namespace keys) |
| PII detection | on (pii_enabled=True) |
always scans content/detail/tags/evidence[]/Assertion.last_evidence on the store path; the configurable pii_action default is warn for the public check_entry_pii helper. On the runtime store path, detected API keys / tokens block the write (PIIBlockError); every other type is recorded in the pii_types metadata and stored verbatim — heuristic detectors do not get to irreversibly rewrite local text. Emails, IPs, SSNs, phone numbers and credit-card shapes are masked at the publish boundary instead. Set pii_custom_patterns to opt in to local masking with your own regexes |
| Poisoning / size-anomaly detection | observe (poisoning_detection_mode="observe") |
the SEC-001 statistical size/tag-count detector records anomaly stats + telemetry but does not quarantine by default; enforce is opt-in. There is no per-source exemption: a caller-supplied metadata['source'] cannot skip enforce-mode quarantine |
| Trust scoring | observe (trust_scoring_mode="observe") |
logs intake trust decisions; enforce/strict are opt-in |
| Provenance signing | required (provenance_required=True) |
persisted rows carry a signed provenance hash-chain |
| Canary tamper response | halt (canary_fail_mode="halt") |
seeded canaries are probed on recall; tamper detection halts by default (degrade/log-only opt-in) |
| Remote sync / publishing | off (sync_enabled=False) |
no remote sync out of the box; local_only=True hard-blocks it via a validator |
| Model remote-code execution | off (embedding_trust_remote_code=False) |
the ONLY input to sentence-transformers' trust_remote_code. Left False, a model repository that ships its own Python modules is refused with RemoteCodeNotPermittedError naming this field; set it true only for a repository you trust, because its code then runs with your process's privileges. The shipped default model needs no remote code, so the secure default is also the working default |
memory.db permissions |
0600 |
the file-backed store is chmod 0600 (owner-only) on creation; a non-POSIX platform degrades to a db_chmod_failed warning |
export TRW_OFFLINE=1 # block the huggingface.co model download (local_files_only)# MemoryConfig
local_only: true # hard-block all remote sync + model downloadFor hybrid recall offline, populate the model cache before enabling either switch, in the same environment: python -c "from sentence_transformers import SentenceTransformer, CrossEncoder; SentenceTransformer('BAAI/bge-small-en-v1.5'); CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')". Otherwise the first embedding load raises LocalOnlyViolationError (an uncached re-ranker is skipped silently and recall keeps fusion order). To run keyword-only without that error, omit the [embeddings] extra entirely. Verify the on-disk memory.db is mode 0600 and that no outbound connection is attempted on first use.
The default embedding model is now BAAI/bge-small-en-v1.5 (same 384 dimensions; queries carry the model's search instruction, stored documents do not). Vectors written by all-MiniLM-L6-v2 — or written before vectors recorded which model produced them — live in a different embedding space, so dense recall ignores them rather than scoring a new-model query against old-model vectors. Until they are re-encoded:
- BM25 still ranks every row, so recall keeps working, with keyword-only relevance for the old rows;
- each recall logs one
dense_vectors_excluded_embedding_spacewarning with the number of vectors it held back.
Re-encode each namespace once (idempotent and resumable — rows already in the active space are skipped, and each batch commits on its own):
trw-memory reembed --namespace default # --batch-size 64, --format jsonasync with MemoryClient(namespace="default") as client:
counts = await client.reembed()The re-embed honours TRW_OFFLINE / HF_HUB_OFFLINE / local_only like every other load: with a switch set and the model not cached it raises LocalOnlyViolationError instead of downloading, so pre-download BAAI/bge-small-en-v1.5 first. To keep the previous model instead, set MEMORY_EMBEDDING_MODEL=all-MiniLM-L6-v2 (or embedding_model in config); it keeps working, without a query instruction. A process only ever scores vectors from the exact space its embedder reports, so switching models back and forth never mixes spaces — it just needs another reembed.
HyPE question generation and HyDE query expansion are removed. Ordinary
embeddings, lexical/hybrid recall, code/wiki references and Distill data are not
removed. Delete imports of QuestionGenerator and NoOpQuestionGenerator.
Remove question_generator, query_expansion, and collapse_hype arguments,
and the hype_enabled, hype_questions_per_entry, hype_min_question_chars
settings (including memory_ aliases and environment/YAML entries).
Explicit neutral legacy settings (False, 3, 8) and API arguments (None
for the generator, None/blank expansion, False for collapse) warn temporarily;
activation, nondefault values and invalid types fail before the operation.
Retired settings no longer appear in emitted configuration. These tombstones
will disappear in the next declared breaking API release after the retirement
release; that release's notes must announce their removal.
Existing derived question vectors are not knowledge records. Recall ignores them
by requiring canonical membership, without excluding real IDs that happen to end
in #hype0. Normal update/forget removes only namespace-owned noncanonical
siblings of the selected canonical parent. Orphan/unknown vector rows remain
untouched for a future canonical-only index rebuild. No startup purge occurs.
Stop old-version writers first; they can regenerate retired vectors. Preserve a
verified backup using SQLite's online backup API (the approach in
storage/_schema_backup.py), not a copy of a live database without its WAL.
Do not overwrite canonical writes made since a snapshot to recover optional
vectors. The retirement itself changes no schema or historical migration.
This recipe is for an existing disposable unencrypted snapshot, not a live
store. Choose its original embedding dimension, namespace and parent IDs
explicitly; opening SQLiteBackend can perform normal schema initialization.
Encrypted stores require their existing key-aware backup/open procedure instead.
apply = False only enumerates selected vectors; changing it to True removes
those derived vectors atomically. It never deletes canonical records or other
namespaces. No vectors installed means unavailable, not a successful cleanup.
from pathlib import Path
from trw_memory.storage.sqlite_backend import SQLiteBackend
snapshot = Path("/absolute/path/to/disposable-snapshot.db")
if not snapshot.is_file():
raise FileNotFoundError(snapshot)
namespace = "default" # explicitly selected, locally authorized namespace
parents = ["selected-parent-id"]
apply = False
backend = SQLiteBackend(snapshot, dim=384) # use this snapshot's dimension
try:
if not backend.supports_vectors():
raise RuntimeError("legacy cleanup unavailable: sqlite-vec required")
with backend.transaction():
for parent_id in parents:
siblings = backend.hype_sibling_ids(parent_id, namespace=namespace)
print(parent_id, siblings)
if apply:
backend.delete_hype_siblings(parent_id, namespace=namespace)
finally:
backend.close()Cleanup is idempotent; interruption rolls the transaction back. Package rollback can reopen the same canonical store; restoring previous optional ranking also requires its matching derived-index snapshot. Never discard newer canonical data for that purpose. Internal cleanup helpers will be removed once the supported store floor rejects pre-retirement stores unless canonical-only vector rebuilding has been verified; ordinary orphan-index handling then owns residual derived data.
trw-memory is tested on CPython 3.10 through 3.14 (this repository's own development
interpreter is CPython 3.14.7). One property of the interpreter matters beyond the version:
its bundled SQLite. WAL space is only RECLAIMED on SQLite >= 3.51.3 (or the 3.44.6 / 3.50.7
backports) — below that, storage/_wal_checkpoint.py coerces resetting checkpoints to
PASSIVE, which is correct and safe but lets the -wal file grow without shrinking. Check
yours with python -c "import sqlite3; print(sqlite3.sqlite_version)"; on macOS, Homebrew's
current Python ships a qualifying build, and trw-mcp doctor names the qualifying
interpreters it finds.
The engine is SELECTED at import by storage/_dbapi.py, which ranks
the interpreter's SQLite against an installed pysqlite3 on (carries the fix, version) and
never replaces a newer engine with an older wheel. The optional [sqlite-fix] extra pulls
pysqlite3-binary on x86_64 Linux only — no published wheel currently bundles a qualifying
SQLite, so it is an engine override, not a fix.
- SQLite driver —
pysqlite3-binaryis no longer a runtime dependency on any platform; it moved to the optional[sqlite-fix]extra, marked for x86_64 Linux (the only platform it publishes a wheel for). It used to be a hard Linux dependency, which made aarch64 Linux installs fail outright while delivering SQLite 3.51.1 — below the 3.51.3 fix it existed to provide. The runtime probe instorage/_dbapi.py, not the dependency name or the package version, decides and reports which engine is active. - Vector search is optional —
[vectors](sqlite-vec) and[embeddings](sentence-transformers) are optional extras. When they are unavailable the retrieval pipeline degrades gracefully to BM25 and/or the backend's built-in keyword search rather than failing.
# Install dev dependencies
pip install -e ".[dev]"
# Run full test suite (>=85% coverage required — see fail_under in pyproject.toml)
python -m pytest tests/ -v --cov=trw_memory --cov-report=term-missing
# Type checking (mypy --strict across the package)
python -m mypy --strict src/trw_memory/
# Targeted testing
python -m pytest tests/test_client_*.py -v
python -m pytest tests/test_retrieval_*.py -v
python -m pytest tests/test_storage_sqlite_*.py -vQuality bar: a broad pytest suite, mypy --strict clean, and a coverage floor of 85% (fail_under in pyproject.toml).
| Extra | Packages | Purpose |
|---|---|---|
[encryption] |
sqlcipher3, keyring, cryptography | Encrypted-at-rest DB (SQLCipher) + key storage |
[embeddings] |
sentence-transformers | Dense vector embeddings (BAAI/bge-small-en-v1.5 by default, 384-dim) |
[vectors] |
sqlite-vec | Vector similarity search in SQLite |
[bm25] |
rank-bm25 | BM25 keyword search |
[all] |
embeddings + vectors + bm25 | The full retrieval stack |
[dev] |
pytest, mypy, ruff, coverage, pip-audit, vulture, deptry | Testing and linting |
There is no [llm] extra and no LLM-backed consolidation. Consolidation
summarises a cluster with a longest-content heuristic; an earlier revision of
this table advertised [llm]/anthropic "LLM-augmented consolidation", which
this package never implemented. The [langchain], [llamaindex], [crewai]
and [all-integrations] extras and their adapter modules were removed as unused
surface — see CHANGELOG.md [Unreleased] Removed.
| Command | Purpose |
|---|---|
trw-memory |
CLI for store/recall/search/forget/consolidate/export/import, plus restore, snapshot (create/list/rotate), wiki-lint, and code-index/code-search/code-symbol |
trw-memory-server |
MCP server (stdio transport) |
A persistent, local-first memory engine for AI agents. It stores memories in SQLite and recalls them with keyword search, or, with the retrieval extras installed, hybrid retrieval (BM25 + dense vectors, fused with Reciprocal Rank Fusion, then a cross-encoder re-ranker). It ships a Python SDK (MemoryClient), a CLI (trw-memory), and an MCP server (trw-memory-server). You do not need TRW Framework to use it.
No. store_conversation() stores every turn verbatim and calls no generative LLM at ingest time; the reader does the inference at recall time. There is also no [llm] extra and no LLM-backed consolidation: consolidation summarises a cluster with a longest-content heuristic. Dense retrieval uses a local sentence-transformers embedding model (BAAI/bge-small-en-v1.5 by default) and a local cross-encoder re-ranker, both from the optional [embeddings] extra.
Yes. With the default configuration all data is local and remote sync is off. TRW_OFFLINE=1 / HF_HUB_OFFLINE=1 or local_only: true force local_files_only=True for the embedding and re-ranker models. If the embedding model is not already cached when one of those switches is set, the first embedding load raises LocalOnlyViolationError: pre-download the model, or omit the [embeddings] extra to run keyword-only. See Telemetry and network behavior.
One small, scoped comparison exists. Using mem0's open-source evaluation suite, unmodified, on LOCOMO conversation 0 (n = 152 questions per system, paired by question, one run each): trw-memory 91.4% [85.9, 94.9] vs mem0 (OSS) 88.2% [82.1, 92.4] at top 10 (McNemar p = 0.38), and 91.4% [85.9, 94.9] vs 92.1% [86.7, 95.4] at top 50 (p = 1.00). Neither test detected a statistically significant difference, which does not establish equivalence or superiority. In that run mem0 made ~2 generative LLM calls per turn to ingest the 419-turn conversation (1 h 28 min); trw-memory made none (~75 s); those figures are ingestion only, not total operating cost. Conditions: a local 8B llama3.1 as answerer, judge and mem0's extraction model, and mem0 run as its open-source SDK, not Mem0 Cloud. See the benchmark section.
Not on a very large namespace. Each namespace contributes at most max(limit * 5, hybrid_search_candidate_pool_size) entries (default 1000) to a search, chosen as the most recently updated rows, so on a larger namespace older entries are not searched and an empty result is not evidence of absence. MEMORY_HYBRID_SEARCH_CANDIDATE_POOL_SIZE widens it at a latency cost. recall(limit=N) can also return fewer than N rows: results the cross-encoder scores below -8 are dropped, though the top min(N, max(5, ceil(N / 2))) are always kept.
In .memory/ relative to the current directory by default (override with MEMORY_STORAGE_PATH), as a local SQLite database per namespace plus an optional YAML sidecar. Nothing leaves the machine unless you enable remote sync.
Yes. Run trw-memory-server (stdio transport), or trw-memory-server serve http for a per-user loopback daemon. See MCP memory server.
[vectors] and [embeddings] are optional extras. When they are unavailable the retrieval pipeline degrades gracefully to BM25 and/or the backend's built-in keyword search rather than failing.
That is an open empirical question. On a controlled recall-dependent benchmark (H1-MEMORY-BENCH), agents with memory solved 58/58 tasks that required a fact from an earlier session and agents without memory solved 0/50 (the fact is absent by construction), which demonstrates the mechanism. Early SWE-bench single-shot runs (n ≥ 40) produced null. See Knowledge compounding, measured.
Business Source License 1.1: source-available, free for non-competing use, converting to Apache 2.0 on 2030-03-21. The package is alpha.
Business Source License 1.1 -- source-available, free for non-competing use. Converts to Apache 2.0 on 2030-03-21.
Built by Tyler Wall · TRW Framework · Documentation · License