clawmem

clawmem is a GitHub-backed conversation and memory plugin for OpenClaw.

What it does:

• Creates one mandatory type:conversation issue per real session and mirrors each user/assistant message as its own transcript comment.
• Before turns: best-effort injects compact context from active type:memory issues when it can do so cheaply.
• On reset/end: best-effort writes the final conversation summary/title. Durable memory retention is skill-driven, not plugin-owned.
• Exposes only operational tools: clawmem_status, clawmem_sync, and clawmem_maintain.

Install

openclaw plugins install @clawmem-ai/clawmem
openclaw plugins enable clawmem
openclaw config set plugins.slots.memory clawmem
openclaw config validate
openclaw gateway restart

After restart, confirm OpenClaw shows ClawMem as the active memory plugin. On first use, clawmem bootstraps each agent identity by calling POST /api/v3/agents on git.clawmem.ai, then writes the returned token plus repo_full_name back into your config under plugins.entries.clawmem.config.agents.<agentId> as that agent's defaultRepo. Automatic flows use that defaultRepo; explicit memory work should resolve the route and operate on GitHub-compatible repos through the bundled skill and gh / gh api. When talking to an older backend that does not expose POST /api/v3/agents, the plugin falls back to the deprecated anonymous bootstrap path.

The package now also ships a bundled clawmem skill for runtime memory behavior:

• GitHub-native recall and retention workflow
• post-install repair and verification guidance
• memory schema and wiki context rules
• concrete GitHub-compatible gh / gh api operations

The website SKILL.md should stay bootstrap-focused. Once the plugin is installed, rely on the bundled plugin skill for day-to-day memory behavior.

Publishing

This repo publishes @clawmem-ai/clawmem through GitHub Actions using npm trusted publishing.

Before the workflow can publish successfully, configure the package on npmjs.com with this trusted publisher:

• Organization or user: clawmem-ai
• Repository: clawmem-openclaw-plugin
• Workflow filename: release.yml

Release flow:

1. Bump package.json to the version you want to ship.
2. Create and push a matching tag such as 0.1.6.
3. GitHub Actions runs .github/workflows/release.yml and publishes with OIDC. No long-lived NPM_TOKEN secret is required.

The workflow intentionally publishes from a tag push instead of workflow_dispatch, because npm validates the workflow filename exactly when using trusted publishing.

Runtime Model

ClawMem is OpenClaw's durable memory system.

• Durable facts, preferences, conventions, decisions, lessons, skill pointers, and active-task state belong in ClawMem memory issues.
• Files remain for tools or humans to read directly.
• Memory routing is per agent identity: plugins.entries.clawmem.config.agents.<agentId>.defaultRepo is the default space, and explicit gh / gh api operations may target other repos.
• Shared or team memory should live in a shared repo, not in one agent's private default repo.
• Use the bundled skill and GitHub-compatible operations for memory work. Plugin tools are operational controls, not memory CRUD wrappers.

Bundled Skill And Docs

The plugin package is now the runtime source of truth:

• Bundled runtime skill: skills/clawmem/SKILL.md
• Runtime references: skills/clawmem/references/
• LoCoMo evaluation harness: scripts/eval/
• Setup/bootstrap guide: the website SKILL.md
• Skill-driven redesign: docs/skill-driven-redesign.md

That bundled skill covers:

• recall and save behavior
• schema discipline and answerable retention
• repo-scoped shared-memory routing
• repair and verification guidance
• raw gh / gh api / curl flows

The LoCoMo harness is the benchmark companion to the runtime plugin. It provisions GitHub-compatible agents/repos, writes normal type:memory issues, recalls only memory issues, runs answer/judge batches, and classifies failures into retention, recall, and answer stages.

If your environment still relies on file-injected reminders such as SOUL.md, AGENTS.md, or TOOLS.md, treat them as optional compatibility snippets rather than the primary runtime source of truth.

Config Reference

Minimal config (after auto-provisioning):

{
  plugins: {
    entries: {
      clawmem: {
        enabled: true,
        config: {
          baseUrl: "https://git.clawmem.ai/api/v3",
          authScheme: "token",
          agents: {
            main: {
              baseUrl: "https://git.clawmem.ai/api/v3",
              defaultRepo: "owner/main-memory",
              token: "<token>",
              authScheme: "token"
            }
          }
        }
      }
    }
  }
}

repo is still accepted as a legacy alias, but new installs should use defaultRepo.

Full config with all options:

{
  plugins: {
    entries: {
      clawmem: {
        enabled: true,
        config: {
          baseUrl: "https://git.clawmem.ai/api/v3",
          authScheme: "token",
          agents: {
            main: {
              baseUrl: "https://git.clawmem.ai/api/v3",
              defaultRepo: "owner/main-memory",
              token: "<token>",
              authScheme: "token"
            },
            coder: {
              defaultRepo: "owner/coder-memory",
              token: "<token>"
            }
          },
          summaryWaitTimeoutMs: 120000,
          memoryAutoRecallLimit: 3,
          memoryAutoRecallStrategy: "query-planner",
          memoryAutoRecallPlannerVariantLimit: 6,
          apiRequestRetries: 3,
          transcriptCommentBatchSize: 1,
          conversationSummaryMode: "llm"
        }
      }
    }
  }
}

Notes

• Conversation comments exclude tool calls, tool results, system messages, and heartbeat noise.
• By default, one normalized user/assistant message maps to one issue comment. transcriptCommentBatchSize exists only for explicit bulk/evaluation runs.
• conversationSummaryMode: "placeholder" skips LLM summary/title generation for bulk or eval runs; keep the default "llm" for normal use.
• Comment writes use stable hidden message markers for retry checks. If mirroring falls behind, finalization waits and clawmem_status surfaces the repair error.
• Each agent_end mirrors conversation comments only; no plugin-owned memory extraction runs after turns.
• Finalization generates only the final conversation issue summary/title.
• Summary failures do not block finalization; the mirrored transcript remains the durable source of truth for manual follow-up.
• Memory search and auto-recall only return open type:memory issues. Closed memory issues are treated as stale.
• ClawMem automatically injects a small set of relevant memories before each turn using the agent's default repo and the backend recall API. Auto-recall is best-effort and quietly skips injection when backend recall is unavailable.
• memoryAutoRecallStrategy: "query-planner" is the default. It runs the full query plus focused compact/core/surface/literal/entity variants concurrently, with duplicate variants removed. memoryAutoRecallPlannerVariantLimit defaults to 6 for recall quality; lower it toward 3 for latency-sensitive deployments that only want full, compact, and core query variants. "single" forces one backend search; "literal-repair" keeps the normal full-query order, but for obvious date/count/name questions it may reserve one low-rank slot for a compact lexical repair hit.
• apiRequestRetries retries transient GitHub-compatible read/search failures such as 429, 5xx, and dropped connections. Write requests are not blindly retried.
• Normal recall is memory-first: type:conversation issues are provenance, audit trail, and rebuild input, not the default serving layer. Important transcript details should be distilled into answerable type:memory issues.
• valid_from / valid_to describe memory validity. Event dates and relative-date conversions that matter for answering belong in the visible ## Memory text.
• Retention should include literal anchor ledgers when needed: dates, durations, quantities, exact item names, and first/last/current/planned facts belong in visible memory text, not only in transcript provenance.
• Retention should also make supported inferences answer-shaped: likely/counterfactual/status/suitability memories need the likely answer, basis, and uncertainty boundary in visible text.
• Always-on ClawMem prompt guidance uses the dedicated memory prompt-registration API on OpenClaw 2026.3.22+. On 2026.3.7 through 2026.3.21, ClawMem falls back to before_prompt_build prependSystemContext. Older hosts still support auto-recall, tools, and conversation mirroring, but they cannot inject the static always-on guidance.
• The plugin exposes clawmem_status, clawmem_sync, and clawmem_maintain only.
• Route resolution is now: agent identity supplies credentials, defaultRepo is the fallback memory space, and explicit GitHub-native operations may override repo per operation.
• Memory issues no longer use session:* labels. Session linkage remains a conversation concern, not part of the durable memory schema.
• Conversation lifecycle is stored in native issue state (open while live, closed after finalize); memory lifecycle uses native issue state too (open active, closed stale).
• Memory issue bodies should use visible GitHub Flavored Markdown plus a minimal hidden <!-- clawmem ... --> metadata block.