Local-first agent memory system with repository-scoped recall, handoffs, and anchored code references
Dejavu provides 10 well-named tools with clear, domain-specific descriptions (avg 95 chars). All tools have descriptions and most parameters are typed. However, input schemas lack formal JSON Schema structure with explicit type declarations and constraints. Parameter descriptions are present but minimal (avg 45 chars). Output schemas are not documented. Error handling guidance is absent. The tools follow a coherent memory/recall pattern but lack the rigor expected for production agent integration.
Expire a slip, regardless of state. Marks the slip as forgotten and removes it from future recalls.
Expire drafts older than 24 hours. Returns count of expired slips. Idempotent.
Close the session with a note for the next agent. Creates or updates the active handoff with summary and next-agent guidance.
Promote draft slips to kept state (survives garbage collection). Optionally triggers auto-rollup into session handoff for chain-shaped slips.
Get the current memory context and state. Returns scope, session, author, and repository information.
Recall slips relevant to a query. Returns ranked hits plus the active handoff for the current session (if any).
Jot a draft slip into memory. Returns the created slip with id, text, and metadata.
Input schemas lack formal JSON Schema structure. Parameters show type hints (e.g., 'number | RecallOptions') but no explicit JSON Schema with $schema, properties, required arrays, or constraints. This prevents LLM-friendly schema validation and discovery.
Output schemas are not documented. Tools return complex objects (e.g., recall returns 'ranked hits plus active handoff') but the response structure, field names, and types are not specified. LLMs cannot plan downstream calls or extract data reliably.
No error handling guidance. Tools lack descriptions of failure modes, recovery steps, or actionable error messages. E.g., 'recall' does not explain what happens if the query is empty or if no slips match.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 43 | 2026-07-28+ | v2 |
Reverse lookup: find what memory is anchored to specific files. Complements recall by answering what code-anchored knowledge exists for files about to be changed.
Record that a recalled slip was helpful. Updates slip trust and relevance signals.
Record that a recalled slip was misleading. Updates slip trust signals negatively.
Parameter descriptions are minimal (avg 45 chars). E.g., 'limitOrOptions' is described as 'Limit on number of results or RecallOptions object...' but does not explain when to use each form, what RecallOptions fields are valid, or constraints on the limit value.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk labels are provided in the evaluation metadata but not exposed in the tool schema. LLMs cannot infer which tools are safe to retry or which modify state.