openclaw-mem gives any agent durable, governed memory:
unified recall across lexical, vector, and graph lanes;
a lifecycle that keeps memories fresh; and receipts for everything —
every inclusion cited, every exclusion
explained, every mutation reversible.
SQLite on your machine. No cloud. No telemetry.
pip install openclaw-context-pack
Long-running agents don't just forget. Their memory degrades in ways similarity search can't fix — and gets worse as retrieval gets better.
Yesterday's truth outranks today's correction, long after it stopped being true.
Tool output, scraped text, injected instructions — hostile content that matches the query slips straight into the prompt.
When the agent goes wrong, nothing says why a memory was included — or what was silently left out.
Capture agent activity into append-only JSONL and a SQLite ledger — categorized into eight deterministic kinds with a six-state lifecycle.
Assemble bounded, cited ContextPack bundles: trust policy filters, category quotas, composite scoring — under an explicit token budget.
Trace receipts explain every include and exclude. Mutations are checkpointed and reversible. The evidence is already on disk.
One entrypoint for agents: openclaw-mem recall "…" routes
lexical / vector / hybrid / graph lanes fail-open, with routing evidence in the receipt.
Architecture → · Command migration →
The reproducible proof runs one query twice against the same synthetic memory. Vanilla packing selects a quarantined row because its text matches. The trust-aware pack excludes it with a written reason — while citation coverage stays intact.
Reproduce it locally{
"synthetic_fixture_only": true,
"quarantined_removed": true,
"citation_coverage_preserved": true,
"trust_policy_explains_exclusion": true
}Quarantined and untrusted records are excluded by policy with written reasons — defense-in-depth against memory poisoning.
Every packed item traces to its source record; citation coverage is measured, not assumed.
Eight bilingual kinds, six lifecycle states, category quotas, use-based decay with protected tiers, reversible soft-archive.
Mutations flow through plan → checkpoint → apply → receipt. DB migration is dry-runnable, backup-first, rollbackable.
"What is currently true about X" — source-linked assertions, timelines, supersession, conflict lint.
Upstream, downstream, lineage, and impact over a SQLite-derived graph — also exposed as read-only MCP tools.
Fixed-seed p95 latencies at 100k rows, committed as SLO receipts — reproduce them from the repo.
| 100k-row lane | target | p95 | |
|---|---|---|---|
| stamped connect | <30 ms | 10.754 ms | ✅ |
| lexical recall | <50 ms | 26.805 ms | ✅ |
| hybrid recall | <200 ms | 75.397 ms | ✅ |
| graph-auto pack | <300 ms | 62.160 ms | ✅ |
| sqlite-vec exact search | <30 ms | 23.030 ms | ✅ |
Default install is SQLite-only; NumPy, sqlite-vec, FastEmbed, and Qdrant are optional lanes that fail open. Details and methodology: v2.0.0 release notes →
Honest framing: mem0, supermemory, and mempalace are excellent at recall benchmarks — openclaw-mem is not trying to beat them at that game. Use them together.
| Recall-focused memory layers | openclaw-mem | |
|---|---|---|
| Primary question | "Did the agent remember the right thing?" | "Should this memory be trusted — and can you prove why it's in the prompt?" |
| Inclusion logic | Similarity scores (opaque) | Explicit receipts with include & exclude reasons |
| Untrusted content | Retrieves whenever it matches | Quarantined by trust policy; exclusion documented |
| Mistake recovery | Delete and hope | Checkpointed mutations with rollback receipts |
| Storage default | Vector DB, often cloud | SQLite + JSONL, local-first |
# a governed memory, from zero
pip install openclaw-context-pack
openclaw-mem init
openclaw-mem store --category preference "Prefers Asia/Taipei for time displays."
openclaw-mem recall "timezone preference" --json
# wire it into your coding agent
openclaw-mem install --harness claude-code
openclaw-mem doctor --harness claude-codeEvery harness, one command. Installers are dry-runnable and atomic, write backups, and verify themselves.
Claude Code · Codex · OpenClaw · Gemini CLI · Cursor · Windsurf · generic
No. The default stack is SQLite + JSONL on your machine. Hybrid recall works locally; there is no hosted service and no telemetry.
Yes — v1.9.x databases stay readable, and migration is explicit: db info to inspect, db migrate --dry-run to preview, backup-first apply with a rollback receipt. Legacy commands remain callable with deprecation guidance.
That's not the axis it competes on. Retrieval is solid (see the SLO and gate receipts above), but the differentiated value is governance: citations, trust policies, receipts, rollback. The reality check stays candid about what is and isn't measured.
Local-first, inspectable, reversible — evaluate it in five minutes with synthetic data, then let it earn its way into your agent stack.
MIT OR Apache-2.0 · built sidecar-first for OpenClaw, at home in any harness · 繁體中文讀者：中文版文件