AGENt memoRy -- Memory infrastructure for AI agents. Provides semantic search and storage of knowledge entries across sessions with support for project scoping, importance scoring, and expiry management.
Agenr demonstrates strong naming conventions (verb_noun pattern: agenr_recall, agenr_store, agenr_extract) and comprehensive parameter descriptions. Tool descriptions are detailed and context-rich, exceeding the 50-200 char baseline for LLM optimization. However, output schemas are not explicitly documented in the visible code, responses are described narratively but lack formal JSON Schema definitions. Input schemas are well-structured with proper type constraints (enums for context, expiry, types; numeric bounds for limit, importance, threshold). Error handling guidance is present in descriptions but not formalized in error response patterns. Security considerations (READ_ONLY vs WRITE risk annotations) are noted but not enforced via tool annotations.
Clear the active project for the current OpenClaw session.
Extract knowledge entries from raw text using LLM analysis. Pass conversation transcripts, documents, meeting notes, or any text and the system will identify and extract durable knowledge entries automatically. Use this when you have a block of text that likely contains multiple pieces of worth-remembering information. The LLM extractor identifies facts, decisions, preferences, and lessons, then structures them as entries. For storing specific, known facts - use agenr_store instead (more precise, no LLM interpretation). Use agenr_extract when the text is long or unstructured and you want the system to figure out what is worth keeping. agenr_extract returns extracted candidate entries. Persist known structured memories with agenr_store.
Show the active project, if any, for the current OpenClaw session.
Retrieve relevant knowledge entries from agenr long-term memory using semantic search. Use mid-session when you need context you do not already have. Session-start recall is already handled automatically.
Retrieve knowledge from the shared memory corpus using semantic search. This is your long-term memory - use it proactively, not as a last resort. WHEN TO RECALL: - Before starting work on any project or topic - pull relevant context first - Before saying "I don't know" or "I don't have that information" - one recall call costs nothing, answering wrong costs trust - When referencing past decisions, preferences, or lessons - do not reconstruct from guesswork - At session start with context='session-start' for bootstrap context QUERY TIPS: - Use specific, focused queries: "Jim's family members" beats "tell me everything" - For personal or non-project queries (family, preferences, identity), use project='*' to search across all projects - personal facts are often stored without project scope - If the first query returns poor results, try rephrasing - different terms surface different entries - Use 'types' filter when you know what you need (for example, types='preference' for user preferences, types='decision' for past choices) PROJECT SCOPING: - Omit 'project' to use the configured default scope (recommended for project-specific work) - Pass project='*' to search everything regardless of project (recommended for personal, cross-cutting, or exploratory queries) - Pass a specific project slug to target one project RECALL SILENTLY when possible - do not narrate "let me check my memory" - just do it and weave results into your response naturally.
Output schemas not formally documented. Tool descriptions explain what is returned narratively (e.g., 'returns extracted candidate entries') but lack JSON Schema definitions for response structure. LLMs cannot reliably parse unstructured output or plan downstream tool chains without explicit response schemas.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) are not present in schema definitions. Risk levels are noted in comments (READ_ONLY, WRITE, REVERSIBLE) but not exposed as formal tool metadata, limiting client-side safety checks and agent planning.
Duplicate tool definitions (agenr_recall appears twice as tools 1 and 6; agenr_store appears twice as tools 2 and 7). This creates ambiguity in tool registration and may confuse LLM tool selection. Consolidate to single canonical definitions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Mark a memory entry as retired (soft delete). Retired entries are permanently excluded from all recall - session-start and explicit queries. Use this when an entry is outdated, superseded, incorrect, or no longer relevant. Prefer retirement over ignoring stale entries - stale entries degrade recall quality for everyone. Set persist=true to make retirement survive database re-ingestion (recommended for entries you definitely want gone permanently). Select entries by entry_id (from recall results) or by exact subject string match.
Explicitly set the active project for the current OpenClaw session. This does not infer a project and only accepts configured project slugs.
Store knowledge entries in the shared memory corpus. This is how you persist decisions, preferences, lessons, and important facts across sessions. Store immediately when you learn something durable - do not wait. WHAT TO STORE: - Decisions and their rationale - User preferences and workflow habits - Lessons learned from mistakes or discoveries - Verified facts about people, projects, and systems - Relationships between people, teams, or concepts - Important events only if they carry forward-looking value WHAT NOT TO STORE: - Routine Q&A that does not carry forward value - Progress snapshots that will be stale in minutes - Information that duplicates what is already in files, git, or issue trackers - Unverified guesses - only store what you know. If uncertain, verify first or note uncertainty in the content IMPORTANCE CALIBRATION: - 10: Rare foundational constraint (almost never used) - 9: Critical durable fact or decision - 8: Significant cross-session context - 7: Normal durable memory (default - most entries belong here) - 6: Routine verified observation - 5: Borderline item - if you are not sure it is worth storing, it probably is not LIFETIME TIERS: - permanent: Kept until explicitly retired. Use for most durable knowledge. - temporary: May auto-expire. Use for short-horizon context (active tasks, temporary states). - core: Injected at every session start. Use VERY sparingly - only for infrastructure facts, hostnames, or universal preferences needed before the first message. If it can be recalled on demand, it should be permanent, not core. WRITE QUALITY: - Content must be self-contained - it should make sense without the conversation that produced it - Include specifics: names, dates, version numbers, file paths. Vague entries are useless on recall. - Subject should be a concise label, not a sentence
Store new knowledge entries in agenr long-term memory. Call immediately after decisions, preferences, lessons, and durable facts - but apply the future-session test first: will a fresh session need this to make a better decision, or are you just logging what happened? Store: architecture decisions, workflow constraints, recurring problems, user preferences, durable technical facts, operational lessons, important open risks. Do not store: version shipping events (changelogs are the record), issue/PR filing records (the tracker is the record), phase plans or release sequencing (stale within a session), progress snapshots (stale within minutes), prompt file locations or build logistics. Do not ask before storing - but do ask whether future-you actually needs it.
Update an existing memory entry's importance or lifetime tier in place. Common uses: - Promote an entry that turned out to be more important than initially scored - Demote an entry that is less relevant than it seemed - Change expiry from temporary to permanent when something proves durably useful - Change expiry from permanent to core only if the entry is truly needed at every session start (very rare) Select entries by entry_id (from recall results) or by exact subject string match.
Session project tools (agenr_set_session_project, agenr_get_session_project, agenr_clear_session_project) have minimal descriptions (under 100 chars). Descriptions lack context on when to use these tools, what 'project scope' means, or how they interact with recall/store operations.
Error handling is descriptive in tool descriptions but not formalized as structured error responses. No evidence of error categorization (retryable vs user-fixable vs fatal) or recovery guidance in error response patterns.