Per-agent memory MCPs — recall + skill-registry + user-model
Context
The OpenClaw 28-agent fleet (per openclaw-vps-architecture) needed persistent, per-agent memory that survives across sessions. Three custom MCP servers provide this: recall (cross-session retrieval), skill-registry (procedural memory), and user-model (self-correcting facts). Together they close "Phase 1" of the agent-memory architecture.
Detail
recall MCP (Phase 1.4)
- What: Per-agent SQLite + FTS5 cross-session recall over briefs / memory / dreams / skills
- 3 tools: recall, reindex, stats
- Index: porter unicode61 tokenizer; auto-reindex on startup if stale (>24h)
- DB:
~/.openclaw/workspace-<agent>/recall.db - Initial index for agent-kom: 35 docs in 123 ms
- Source: VPS
~/mcp-servers/recall-mcp/, GitLab project 243
skill-registry MCP (Phase 1.6)
- What: Per-agent procedural memory — markdown skill files with YAML frontmatter
- 5 tools: list, get, save, remove, record_run
- Storage:
~/.openclaw/workspace-<agent>/skills/<name>.md - Naming: DNS-style names with path-traversal guards, 10KB body cap
- Audit log:
~/.openclaw/logs/skill-registry.jsonl - Source: GitLab project 245. Closes Phase 1.
user-model MCP (Phase 1.5)
- What: Lets agent-kom self-persist correction facts to its own MEMORY.md DYNAMIC section
- 3 tools: read, update, remove
- Limits: 200-char fact cap, 2200-char file cap with FIFO eviction
- Audit log:
~/.openclaw/logs/user-model-updates.jsonl - Source: VPS
~/mcp-servers/user-model-mcp/
How they relate to Knowledge OS
These are agent-private memory — distinct from Knowledge OS (which is shared org knowledge):
| Layer | Scope | Content | MCP |
|---|---|---|---|
| recall | Per-agent | That agent's briefs/memory/dreams/skills | recall-<id> |
| skill-registry | Per-agent | That agent's learned procedures | skill-registry-<id> |
| user-model | Per-agent (agent-kom) | Self-correction facts | user-model |
| Knowledge OS | Org-wide, shared | Suppliers, decisions, projects, people | ko-knowledge |
The key distinction: recall/skill-registry are siloed per agent (agent-praew can't see agent-kom's recall). Knowledge OS is the shared layer all agents query. This is WHY KO exists — before it, agents hallucinated org-Q&A because per-agent recall only sees per-agent content.
Per-agent MCP rollout (2026-05-04, team-wide)
- recall + skill-registry extended from agent-kom-only to all 8 active team agents
- Each agent: own
recall-<id>andskill-registry-<id>server instance bound viaOPENCLAW_AGENT_IDenv - Naming: agent-kom unsuffixed (legacy), all others suffixed
Backup
- All 28 agent MEMORY.md + agent-kom skills backed up at GitLab
openclaw-agent-memories(project 246) - Not auto-synced; push manually after material edits
The AIOS Wiki gap (Phase 2.1 — SPEC ONLY)
A shared markdown wiki at GitLab aspire/aios-wiki, mounted on OpenClaw, indexed by recall via a new WIKI_DIR env, would unblock org-knowledge questions for ALL agents. Currently agents hallucinate org Q&A because per-agent recall only sees per-agent content. KO Stage 1 + this AIOS Wiki idea overlap — worth reconciling which is the canonical shared layer (KO vault is the stronger candidate now that it has 40+ pages).
Related
- openclaw-vps-architecture — where these MCPs run
- knowledge-os-stage-1 — the shared-knowledge layer (vs per-agent memory)
- aspire-mcp-ecosystem — full MCP tier matrix