Total recall. 90% fewer tokens.
Your agent forgets past decisions and burns tokens re-reading the same context. Memory Stack runs 5 search engines locally, returns only what matters, and never loses a fact.
> Free and open source. MIT license. No subscription, no cloud API costs.
> All search and storage runs entirely on your machine.
> GitHub: github.com/Apptah/openclaw-memory-stack
┌──────────────────────────────────────────────────────────────┐
│ OPENCLAW MEMORY STACK │
├──────────────────────────────────────────────────────────────┤
│ │
│ SEARCH PIPELINE (runs on every conversation turn) │
│ │
│ ┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐ │
│ │ E1 │ │ E2 │ │ E3 │ │ E4 │ │ E5 │ │
│ │ Full │ │Vector│ │ DAG │ │ Fact │ │ MD │ │
│ │ Text │ │Search│ │Compr.│ │Store │ │Files │ │
│ └──┬───┘ └──┬───┘ └──┬───┘ └──┬───┘ └──┬───┘ │
│ └────────┴────────┴────────┴────────┘ │
│ │ │
│ ▼ │
│ ┌──────────────┐ │
│ │ Result Fusion│ │
│ │ + Reranking │ │
│ └──────────────┘ │
│ │ │
│ ┌─────────┼─────────┐ │
│ ▼ ▼ ▼ │
│ ┌──────┐ ┌──────┐ ┌──────┐ │
│ │ L0 │ │ L1 │ │ L2 │ Token Budget │
│ │~100t │ │~800t │ │ full │ Control │
│ └──────┘ └──────┘ └──────┘ │
│ │
│ CAPTURE (runs after every conversation turn) │
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Fact │ │ Entity │ │ Dedup & │ │
│ │ Extraction │ │ Tracking │ │ Supersede │ │
│ │ (8 types) │ │ (queryable) │ │ (3-level) │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
│ │
│ CROSS-AGENT SHARING │
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ CLI API │ │ Drop Zone │ │ Unified │ │
│ │ query/add/ │ │ ~/.openclaw │ │ Global │ │
│ │ recent │ │ /external/ │ │ Memory │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
│ ▲ ▲ ▲ │
│ Claude Code Cursor / Windsurf Any MCP client │
│ │
│ SELF-HEALING MAINTENANCE (24h cycle, zero config) │
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Auto-Init │ │ Graceful │ │ Health │ │
│ │ workspace │ │ Fallback │ │ Monitor │ │
│ │ + indexing │ │ FTS5-only │ │ + alerting │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
│ │
└──────────────────────────────────────────────────────────────┘
Find function names, variable names, or any pattern across your entire memory — instantly. No extra tools needed, works offline, gets faster the more you use it.
Five engines search in parallel on every conversation turn:
| Engine | What it does |
|---|---|
| -------- | ------------- |
| Full-text | Keyword matching with relevance ranking |
| Vector | Semantic search — understands meaning, not just words |
| Compressed history | Conversation DAG with drill-down |
| Fact store | Structured facts — decisions, deadlines, requirements |
| Markdown | Scans memory files directly |
Results are merged with rank fusion, deduplicated, reranked for diversity, and scored with time decay.
Every wasted token is money burned. Memory Stack eliminates the waste.
| Tier | Tokens | When used |
|---|---|---|
| ------ | -------- | ----------- |
| L0 | ~100 | Auto-recall every turn — minimal cost |
| L1 | ~800 | On-demand search summary |
| L2 | full | Full content on request |
The agent generates summaries — no separate LLM service required.
When the OpenClaw agent stores or surfaces a memory, it should also produce L0/L1
summaries from its own context (it is already an LLM) and persist them via:
openclaw-memory tier set <file> \
--l0 "<one-sentence abstract, ≤100 tokens>" \
--l1 "<focused overview, ≤800 tokens, covers what + why + when to use>"
Read back at any tier:
openclaw-memory tier show <file> --level L0 # or L1, L2
For unattended/offline contexts where no agent is available, fall back to a pure
heuristic (first non-empty line for L0, char-bounded prefix for L1):
openclaw-memory tier auto <file>
Extracts 8 fact types automatically: decisions, deadlines, requirements, entities, preferences, workflows, relationships, corrections.
Works across Claude Code, Cursor, Windsurf, and any MCP-compatible client.
openclaw-memory query/add/recent — any tool can read/write facts.md files into ~/.openclaw/memory/external/ — auto-ingested, no duplicates~/.openclaw/memory/ — your memory follows you across workspacesZero maintenance. The plugin takes care of itself.
| Native | Memory Stack | |
|---|---|---|
| --- | -------- | ------------- |
| Search engines | 2 | 5 (parallel, fused) |
| Token efficiency | Full text every time | Up to 90% fewer |
| Output tiers | Full text | 3 tiers (~100 / ~800 / full) |
| Fact extraction | No | 8 types, structured, negation-aware |
| Duplicate handling | Can pay twice | 3-level dedup, auto-supersede |
| Entity tracking | No | Yes, queryable |
| Cross-agent | No | CLI + drop zone, works with any tool |
| Memory across projects | Separate per workspace | Unified global memory — follows you everywhere |
| Self-healing | No | Auto-maintain, auto-fallback |
| Runs locally | Yes | Yes — all search runs locally, no cloud dependency |
Most AI memory systems make you choose between control and convenience. Memory Stack doesn't.
| Hermes Agent | Google Always-On | Memory Stack | |
|---|---|---|---|
| --- | --- | --- | --- |
| Trigger | Explicit tool call | Automatic (Google decides) | Deterministic router — 7 explicit rules |
| Routing logic | Agent judgment (opaque) | Black box | Keyword pattern matching — fully inspectable |
| Fallback on poor recall | None | Uncontrollable | Sequential chain — relevance < 0.4 → next backend |
| Storage location | Developer-managed DB | Google cloud | Local git branch + SQLite — nothing leaves your machine |
| Token control | None | None | 3 tiers: ~100 / ~800 / full |
| Deduplication | Manual | Unknown | 4-level auto-dedup |
| Works offline | Depends on DB | No | Yes — core search fully offline |
| Open source | Varies | No | MIT — full source on GitHub |
| Setup time | Hours | Minutes (Gemini only) | One command, any MCP client |
Hermes gives developers control but no routing or fallback infrastructure. Google gives convenience but no privacy and no control. Memory Stack gives you both.
npx clawhub@latest package install openclaw-memory-stack
Or clone and run directly:
git clone https://github.com/Apptah/openclaw-memory-stack
cd openclaw-memory-stack
./install.sh
Runs on macOS, Linux, and Windows (WSL2). Requires bash, python3, and OpenClaw 2026.3.2 or later. Bun is optional (enables QMD vector search).
The included install.sh performs the following actions:
~/.openclaw/memory-stack/, registers plugin in ~/.openclaw/extensions/openclaw-memory-stack/, updates ~/.openclaw/openclaw.jsoninstall.sh --upgrade downloads new versions with mandatory SHA-256 checksum verificationOPENCLAW_LLM_API_KEY / OPENAI_API_KEY env vars are optional, used only for LLM-powered fact extraction. If not set, falls back to local Ollama/MLX. These keys are sent only to the endpoint you configure (default: api.openai.com), never to our serversAll search and storage runs entirely on your machine.
MIT. Free and open source. See LICENSE.
Questions or issues? Open a GitHub issue at github.com/Apptah/openclaw-memory-stack.
共 1 个版本