Files
openclaw/docs/concepts/memory.md
2026-07-28 22:26:19 -04:00

14 KiB

summary, title, read_when
summary title read_when
How OpenClaw remembers things across sessions Memory overview
You want to understand how memory works
You want to know what memory files to write

OpenClaw remembers things by writing plain Markdown files in your agent's workspace (default ~/.openclaw/workspace). The model only remembers what gets saved to disk; there is no hidden state.

How it works

Your agent has four memory-related files:

  • USER.md (optional) — stable preferences, communication style, relationships, and active-project context written as directives. Loaded at the start of a session with a separate small budget.
  • MEMORY.md — long-term memory. Durable non-profile facts and decisions. Loaded at the start of a session.
  • memory/YYYY-MM-DD.md (or memory/YYYY-MM-DD-<slug>.md) — daily notes. Running context and observations. Today's and yesterday's dated notes load automatically on a bare /new or /reset; slugged variants, such as those written by the bundled session-memory hook, are picked up alongside the date-only file.
  • DREAMS.md (optional) — Dream Diary and dreaming sweep summaries for human review, including grounded historical backfill entries.
If you want your agent to remember something, just ask it: "Remember that I prefer TypeScript." It writes the note to the appropriate file.

What goes where

USER.md is the compact user-model layer. Write stable preferences and profile facts as imperative directives with observed-date and active/superseded metadata. When a preference changes, supersede it in place instead of appending a contradictory active directive. See User model.

MEMORY.md is the compact, curated layer for durable non-profile facts, standing decisions, and short summaries that should be available at the start of a session. It is not a raw transcript, daily log, or exhaustive archive.

memory/YYYY-MM-DD.md files are the working layer: detailed daily notes, observations, session summaries, and raw context that may still be useful later. These are indexed for memory_search and memory_get, but are not injected into the bootstrap prompt on every turn.

Over time, useful material from daily notes is distilled into MEMORY.md by the default dreaming sweep. The generated workspace instructions still encourage the agent to record durable facts as it works, while dreaming handles background consolidation. The default heartbeat prompt performs no memory maintenance on its own.

If MEMORY.md grows past the bootstrap file budget, OpenClaw keeps the file on disk intact but truncates the copy injected into context. Treat that as a signal to move detailed material into memory/*.md, keep only a durable summary in MEMORY.md, or raise the bootstrap limits if you want to spend more prompt budget. Use /context list, /context detail, or openclaw doctor to see raw vs. injected sizes and truncation status.

Import from coding assistants

The Control UI can import existing local memory from Codex, Claude Code, and Hermes. Open SettingsImport Memory, choose the destination agent, review the detected files, and confirm the import. For the existing default agent, you can instead open Settings → Ask OpenClaw and say import memory; this narrower chat wizard requires completed onboarding, copies only new detected memory, and reports per-source failures or possible partial copies. OpenClaw copies only Markdown memory:

  • Codex: the consolidated MEMORY.md and memory_summary.md files under ~/.codex/memories (or CODEX_HOME/memories). Raw rollout and transcript files are not imported.
  • Claude Code: Markdown files from each project auto-memory directory under ~/.claude/projects/*/memory, plus a user-configured autoMemoryDirectory when present. Project instructions, sessions, settings, and credentials are not part of this memory-only action.
  • Hermes: MEMORY.md and USER.md from the detected Hermes home. Config, credentials, and skills are not part of this memory-only action.

Imported files stay separate under memory/imports/codex/ and memory/imports/claude-code/, or memory/imports/hermes/ in the selected agent workspace. They are indexed for memory_search and available through memory_get; they are not merged into the agent's bootstrap MEMORY.md. The source files are left unchanged.

The preview marks destination conflicts. Enable Replace existing imports to replace those files; apply creates a verified pre-import backup and preserves item-level copies of overwritten files in the migration report.

Action-sensitive memories

Most memories are ordinary Markdown notes. Some affect what the agent should do later; for those, capture when it is safe to act on the note, not just the fact itself.

Capture that action boundary when a note involves:

  • approval or permission requirements,
  • temporary constraints,
  • handoffs to another session, thread, or person,
  • expiry conditions,
  • safe-to-act timing,
  • source or owner authority,
  • instructions to avoid a tempting action.

A useful action-sensitive memory makes clear:

  • what changes future behavior,
  • when or under what condition it applies,
  • when it expires, or what unlocks action,
  • what the agent should avoid doing,
  • who is the source or owner, if that affects trust or authority.

Memory can preserve approval context, but it does not enforce policy. Use OpenClaw approval settings, sandboxing, and scheduled tasks for hard operational controls.

Example:

The API migration is being designed in another session. Future turns should
not edit the API implementation from this thread; use findings here only as
design input until the migration plan lands.

Another example:

A report from an untrusted source needs review before promotion. Future turns
should treat it as evidence only; do not store it as durable memory until a
trusted reviewer confirms the contents.

This is not a required schema for every memory; simple facts can stay concise. Use action-sensitive boundaries when losing timing, authority, expiry, or safe-to-act context could cause the agent to do the wrong thing later.

Use scheduled tasks for exact reminders, timed checks, and recurring work. Memory can still summarize the durable context around that work.

Retired inferred commitments

Some future follow-ups are not durable facts. If a future event should trigger an action, use a standing intent. If a clock time should trigger it, use a scheduled task.

The inferred commitments experiment is retired. OpenClaw no longer extracts or delivers those follow-ups. Use scheduled tasks for future actions; the legacy openclaw commitments command remains available to inspect or dismiss existing stored rows.

Memory tools

The agent has three tools for working with memory:

  • memory_search — finds relevant notes using semantic search, even when the wording differs from the original.
  • memory_get — reads a specific memory file or line range.
  • intent — creates, lists, or explicitly cancels event-conditioned standing intents. Time-based reminders continue to use scheduled tasks.

Both tools are provided by the active memory plugin (default: memory-core).

When an embedding provider is configured, memory_search uses hybrid search: vector similarity (semantic meaning) combined with keyword matching (exact terms like IDs and code symbols). This works out of the box with an API key for any supported provider.

OpenClaw uses OpenAI embeddings by default. Set `memory.search.provider` explicitly to use Gemini, Voyage, Mistral, Bedrock, DeepInfra, local GGUF, Ollama, LM Studio, GitHub Copilot, or a generic OpenAI-compatible endpoint.

See Memory search for how search works, tuning options, and provider setup.

Memory backends

SQLite-based. Works out of the box with keyword search, vector similarity, and hybrid search. No extra dependencies. Local-first sidecar with reranking, query expansion, and the ability to index directories outside the workspace. AI-native cross-session memory with user modeling, semantic search, and multi-agent awareness. Plugin install. LanceDB-backed memory with OpenAI-compatible embeddings, auto-recall, auto-capture, and local Ollama embedding support. Plugin install.

Knowledge wiki layer

If you want durable memory to behave more like a maintained knowledge base than raw notes, use the bundled memory-wiki plugin. It compiles durable knowledge into a wiki vault with deterministic page structure, structured claims and evidence, contradiction and freshness tracking, generated dashboards, compiled digests, and wiki-native tools (wiki_status, wiki_search, wiki_get, wiki_apply, wiki_lint).

memory-wiki does not replace the active memory plugin; the active memory plugin still owns recall, promotion, and dreaming. memory-wiki adds a provenance-rich knowledge layer beside it.

Compiles durable memory into a provenance-rich wiki vault with claims, dashboards, bridge mode, and Obsidian-friendly workflows.

Automatic memory flush

Before compaction summarizes your conversation, OpenClaw runs a silent turn that reminds the agent to save important context to memory files. This is on by default; set agents.defaults.compaction.memoryFlush.enabled: false to turn it off.

To keep that housekeeping turn on a local model, set an exact override that applies only to the memory-flush turn (it does not inherit the active session's model fallback chain):

{
  "agents": {
    "defaults": {
      "compaction": {
        "memoryFlush": {
          "model": "ollama/qwen3:8b"
        }
      }
    }
  }
}
The memory flush prevents context loss during compaction. If your agent has important facts in the conversation that are not yet written to a file, they are saved automatically before the summary happens.

Dreaming

Dreaming is the default background consolidation path for memory. It collects short-term recall signals, scores candidates, and promotes only qualified owner or agent-derived items into long-term memory (MEMORY.md):

  • Default on: disable it with plugins.entries.memory-core.config.dreaming.enabled: false.
  • Scheduled: when enabled, memory-core auto-manages one recurring cron job for a full dreaming sweep.
  • Thresholded: promotions must pass score, recall-frequency, and query-diversity gates.
  • Consolidated: a bounded subagent rewrite merges duplicates and supersedes stale entries after the deterministic gate. Invalid or unavailable rewrites use append-only fallback.
  • Taint gated: untrusted and system-derived candidates never enter the consolidation prompt or durable promotion path.
  • Reviewable: phase summaries and diary entries are written to DREAMS.md for human review, including rewrite counts and highlights.

This background pattern follows the motivation behind sleep-time compute (arXiv:2504.13171). Provenance-aware reflection also follows the durable memory lessons of the Generative Agents research.

See Dreaming for phase behavior, scoring signals, and Dream Diary details.

Grounded backfill and live promotion

The dreaming system has two related review lanes:

  • Live dreaming works from the short-term dreaming store under memory/.dreams/ and is what the normal deep phase uses to decide what graduates into MEMORY.md.
  • Grounded backfill reads historical memory/YYYY-MM-DD.md notes as standalone day files and writes structured review output into DREAMS.md.

Grounded backfill is useful for replaying older notes and inspecting what the system considers durable, without manually editing MEMORY.md.

openclaw memory rem-backfill --path ./memory --stage-short-term

The --stage-short-term flag stages grounded durable candidates into the same short-term dreaming store the normal deep phase already uses; it does not promote them directly. So:

  • DREAMS.md stays the human review surface.
  • The short-term store stays the machine-facing ranking surface.
  • MEMORY.md is still only written by deep promotion.

To undo a replay without touching ordinary diary entries or normal recall state:

openclaw memory rem-backfill --rollback
openclaw memory rem-backfill --rollback-short-term

CLI

openclaw memory status          # Check index status and provider
openclaw memory search "query"  # Search from the command line
openclaw memory index --force   # Rebuild the index

Further reading