Memory plugin for OpenClaw agents.

Formative Memory is an OpenClaw plugin that gives your agent long-term memory modeled after how biological memory works. Before every response, it recalls relevant memories into context. After every response, it evaluates which memories actually contributed — strengthening useful ones and letting unused ones fade. Every night, the agent sleeps: a consolidation process decays, prunes, merges, and connects memories to keep recalled context high-quality.

Over time, raw facts combine into richer structures: merged summaries, connected associations, and deeper understanding.

How It Works

Recall

Before every response, the plugin searches for memories relevant to the current conversation using hybrid search (embedding similarity + BM25 full-text), ranked by memory strength. Strongly-associated neighbors of the top results are also pulled in through single-hop association expansion, so related memories surface even when they don't directly match the query. Matching memories are injected into the agent's context automatically. Each retrieval strengthens the memories that were surfaced.

User: "Do you remember that restaurant? The one by the beach
       last summer. I'm trying to book for our anniversary."

Injected memories:
  [a3f2|fact|strength=0.82]  "Dinner at Maininki, Hanko, April 2026"
  [b7d1|event|2026-07-04]    "Wedding anniversary — 12 years"
  [c9f3|preference|str=0.74] "Sanna loves peonies"
  [d2e6|fact|2026-07-02]     "Sanna: private doctor's appointment"

Agent: "Of course! It was Maininki, in Hanko. Shall I book a table?
        Your 12th anniversary is coming up on July 4th."

The agent sees recalled memories as context, not instructions — this reduces prompt injection risk from stored content.

Evaluate

After each response, the plugin tracks which memories were surfaced and whether they actually influenced the reply. This happens at two levels:

• Automatic attribution — the plugin logs which memories were injected and which the model referenced, building a retrieval history without any agent effort
• Explicit feedback — the agent can call memory_feedback to rate a memory's usefulness (1–5), signaling quality directly

Both signals feed into consolidation: frequently used, highly rated memories are reinforced, while memories that are surfaced but never referenced gradually lose strength. This creates a feedback loop where the memory system learns what is actually useful, not just what matches a query.

After the agent's response about Maininki:

Automatic attribution (logged by the plugin):
  ✓ [a3f2] "Dinner at Maininki, Hanko"     — referenced in reply
  ✓ [b7d1] "Wedding anniversary — 12 years" — referenced in reply
  · [c9f3] "Sanna loves peonies"            — injected, not used
  · [d2e6] "Sanna: doctor's appointment"    — injected, not used

  → a3f2 and b7d1 are reinforced at next consolidation
  → c9f3 and d2e6 were surfaced but ignored — no reinforcement

Explicit feedback (agent calls memory_feedback):
  → memory_feedback(memory_id: "a3f2", rating: 5)
  → "Directly answered the user's question"

Capture

Memories are collected in two ways. The agent can store a memory explicitly with memory_store, and auto-capture extracts durable facts from conversations automatically after each turn. Extraction uses chain-of-thought reasoning — the LLM evaluates each candidate fact for durability beyond the current task, discarding ephemeral details like "currently looking for a birthday gift".

After the turn above, auto-capture extracts:
→ store("Booking anniversary dinner at Maininki", type: event,
        temporal_anchor: 2026-07-04, temporal_state: future)

A later turn — user asks for Sanna's favorite foods:
Agent explicitly stores:
→ memory_store("Sanna's favorites: salmon soup (her mother's recipe,
   no cream), pistachio ice cream, meat pies from Market Hall
   on Saturdays", type: preference)
→ id: e8b2a1f4, strength: 1.0

Each memory is content-addressed (SHA-256) — same content always produces the same ID, so duplicates are prevented by design.

Consolidate

Every night, the agent sleeps. A consolidation process runs through the accumulated memories:

Before consolidation:
  [a3f2|strength=0.82] "Dinner at Maininki, Hanko, April 2026"
  [f1c4|strength=0.65] "Maininki — beachfront restaurant, good wine list"
  [a9b3|strength=0.41] "Tried booking Maininki in June, fully booked"

After consolidation:
  [g7e2|strength=1.00] "Maininki, Hanko: beachfront restaurant with good
   wine list. Visited April 2026. Book early — fills up in summer."

  Associations formed:
    "Maininki" ←0.7→ "Wedding anniversary"
    "Maininki" ←0.4→ "Sanna loves peonies"

All mutation happens during consolidation — live chat stays fast and predictable. Over time, simple facts combine into richer structures: merged summaries, connected associations, and deeper understanding.

Quick Start

Install the plugin:

openclaw plugins install formative-memory

This installs from npm, enables the plugin, and assigns it the memory slot automatically. Restart the gateway to load the plugin.

That's it. The plugin works out of the box:

• Auto-capture records conversations for consolidation
• Auto-recall surfaces relevant memories before every response
• Consolidation runs automatically to maintain memory quality
• Startup tasks (migrating existing memory files, scrubbing legacy memory instructions from AGENTS.md, backfilling embeddings) run automatically at gateway boot

No configuration needed — sensible defaults are built in.

Security: plugin allowlist. OpenClaw logs a warning if plugins.allow is empty — non-bundled plugins auto-load without validation. To silence it and restrict loading to trusted plugins, add an explicit allowlist in openclaw.json:

{
  "plugins": {
    "allow": ["formative-memory"]
  }
}

Multi-agent setups: API key resolution is delegated to the OpenClaw SDK's resolveApiKeyForProvider, which handles auth profiles, multi-agent resolution, and credential precedence internally. The plugin resolves keys lazily — if the runtime context (agentDir) is not yet available at boot (e.g. during heartbeat), resolution defers until the first tool call or service start provides it.

Memory Tools

The plugin registers five tools the agent can use during conversation:

Memory types: fact, preference, decision, plan, observation.

Configuration

All settings are optional — defaults are designed to work out of the box. Configuration goes in openclaw.json under the plugin entry:

{
  "plugins": {
    "entries": {
      "formative-memory": {
        "enabled": true,
        "config": {
          "autoRecall": true,
          "autoCapture": true,
          "requireEmbedding": true,
          "embedding": {
            "provider": "auto"
          },
          "consolidation": {
            "notification": "errors",
            "errorNotification": true
          },
          "temporal": {
            "notification": "errors",
            "errorNotification": true
          }
        }
      }
    }
  }
}

The "auto" provider selects the best available embedding provider from your configured API keys. When requireEmbedding is true (the default), the plugin will not start without a working embedding provider. Set it to false to allow graceful degradation to keyword-only search.

We strongly recommend configuring an OpenAI or Google API key for embeddings. Without embeddings, memory search operates in keyword-only (BM25) mode which cannot match paraphrases, synonyms, or typos. For example, searching for "shipping date" will not find a memory about "release deadline". Embeddings dramatically improve recall quality.

Salience Profile

A default salience.md is created automatically in the memory directory (default: ~/.openclaw/memory/associative/salience.md) if one does not already exist. This file guides both auto-capture extraction and the agent's memory_store decisions — edit it to match your priorities.

Write it in natural language — describe what kinds of information matter to you. For example:

Pay attention to:
- Family members and relationships
- Health details I share
- Travel plans and locations
- Professional milestones and career goals

Less important:
- What I'm currently shopping for or browsing
- Passing mentions of things I'm not committed to

The profile is capped at 4000 characters and treated as preference guidance only — it cannot override the plugin's extraction rules or output format.

API Keys

API key resolution is delegated to the OpenClaw SDK's resolveApiKeyForProvider, which handles auth profiles, env vars, OAuth, and multi-agent resolution internally. Configure a profile under the standard OpenClaw setup:

{
  "version": 1,
  "profiles": {
    "openai:default": { "type": "api_key", "key": "sk-..." },
    "google:default": { "type": "api_key", "key": "AIza..." }
  }
}

The openai:default and google:default profile names are picked up automatically. If you have multiple profiles for the same provider (e.g. openai:work and openai:personal), the plugin warns and picks the first one — add a :default profile to select explicitly.

Provider pinning

The plugin pins the selected provider and model to the database on first successful resolution. On subsequent runs, the same provider and model are used regardless of embedding.provider in config — this prevents silent drift that would corrupt the vector store when a different provider (producing different-dimension vectors) takes over.

If you intentionally want to switch providers or models for an existing database, you must re-embed all memories via migration. Attempting to change the configured provider mid-life produces a clear error at startup rather than silent corruption.

Disabling Overlapping Memory Features

OpenClaw ships with built-in memory features that overlap with this plugin. We recommend disabling them to avoid redundant injection and conflicting memory writes:

{
  "plugins": {
    "entries": {
      "active-memory": { "enabled": false },
      "memory-core": { "enabled": false }
    }
  },
  "hooks": {
    "session-memory": { "enabled": false }
  }
}

If Active Memory is left enabled, the plugin logs a warning at startup and automatically reduces its own recall limits to minimize redundancy. Disabling is the cleaner approach.

Architecture

OpenClaw Runtime
    |
    |-- Context Engine --- assemble() -> association-augmented recall
    |   (budget-aware)     afterTurn() -> provenance + auto-capture
    |                      cancelExtractions() -> abort in-flight tasks
    |                      dispose() -> await tasks, reset caches
    |
    |-- Memory Tools ----- memory_store    - create memory
    |   (agent-initiated)  memory_search   - hybrid search
    |                      memory_get      - lookup by ID
    |                      memory_feedback - rate usefulness
    |                      memory_browse   - browse by importance
    |
    +-- Consolidation ---- reinforce -> decay -> associate -> transition
        (automatic)        -> prune -> merge (LLM) -> cleanup

Storage: SQLite with FTS5 for full-text search. Single file, no external services.

Embedding: Standalone fetch-based clients for OpenAI and Gemini read keys from auth-profiles.json. Additional providers (Voyage, Mistral, Ollama, local) resolve through memory-core's embedding adapter registry when installed. Circuit breaker with graceful fallback to keyword-only search when requireEmbedding is false.

Consolidation LLM: Uses Anthropic (Claude) or OpenAI for memory merging. Runs only during consolidation, not during normal chat.

Trust & Security

Automatically recalled memories are framed as reference data, not instructions. This reduces prompt injection risk from stored memory content, but memory remains untrusted input — the framing is probabilistic, not a hard security boundary.

Do not store secrets (API keys, passwords) in memories. They will be surfaced to the model during recall.

Privacy

This plugin sends conversation text to the LLM and embedding providers you configure in OpenClaw. Be aware of what leaves your machine:

• Auto-capture (autoCapture, default true) — at the end of each agent turn, recent user/assistant text is sent to the configured LLM provider for memory extraction.
• Embeddings — every stored memory is sent to your configured embedding provider (OpenAI / Google / etc.) to produce vectors. New recall queries are also embedded.
• Consolidation (nightly "sleep") — candidate memory pairs are sent to the configured LLM provider to be merged or rephrased.
• Query logging (logQueries, default false) — when enabled, raw query text appears in OpenClaw logs.

What stays local: the SQLite database (at dbPath, default ~/.openclaw/memory/associative), all memory content at rest, and BM25 search indices.

To minimize external exposure:

• Set autoCapture: false to require explicit memory_store calls.
• Set requireEmbedding: false to fall back to BM25-only search and avoid sending text to an embedding provider altogether.
• Choose a provider whose data-retention policy you trust, or run a local OpenAI-compatible endpoint.
• Keep logQueries: false (the default) if logs may be shared.

CLI

A standalone diagnostic CLI operates directly on the SQLite database — no OpenClaw runtime needed:

memory stats <memory-dir>         # Database overview
memory list <memory-dir>          # List memories (filterable)
memory inspect <memory-dir> <id>  # Detailed view of a single memory
memory search <memory-dir> <q>    # Search by content
memory export <memory-dir>        # Export to JSON
memory history <memory-dir>       # Retrieval history
memory graph <memory-dir>         # Association graph

Logging

Centralized logging with configurable verbosity. By default only significant events are logged (info level).

Enable debug logging:

{ "verbose": true }

All log lines are prefixed with [formative-memory] [level] for easy filtering. Query text is never included in logs by default — set logQueries: true to opt in.

Documentation

• How Formative Memory Works — conceptual guide
• Architecture — storage, retrieval, provenance, consolidation
• OpenClaw Memory Systems — comparison and coexistence with built-in memory
• Glossary — terminology

Development

pnpm install          # Install dependencies
pnpm build            # Build (tsdown)
pnpm test             # Run tests (vitest)
pnpm lint             # Lint (oxlint)
pnpm check            # Full check (format + typecheck + lint)

Requires Node.js >= 22.12.0, pnpm 10.x.

Contributing

Contributions welcome. Areas where help is especially useful:

• Consolidation algorithm tuning and evaluation
• Embedding model benchmarks
• Adapters for other AI coding agents
• Documentation and examples

License

MIT