Layered persistent memory for AI agents — hybrid search (BM25 + vector) with SQLite backend, encryption, and multi-adapter support (MCP, HTTP, pi-extension)
Memnest provides 6 tools with complete schemas and descriptions. Naming follows verb_noun convention (memory_*, secret_*). Descriptions are detailed and context-rich (100-300 chars), exceeding baseline averages. However, parameter descriptions lack actionable constraints (no enums for importance/memory_kind, no min/max for n_results/max_chars despite having them in schema). Error handling is minimal, no recovery guidance or categorization. Output schemas are undocumented. The server is STDIO-only, which caps protocol readiness at 50 regardless of definition quality.
Delete a memory by ID. Deletion is permanent and cannot be undone.
Read a redacted memory page and optional same-session transcript neighbors in capture order. max_chars caps total document text; fetch clipped neighbors individually by id.
Save something the next session should still know. Call it without being asked when the user corrects you, states a preference or a decision, or when you learn a config value, port, path, or a fix that took real effort to find. After changing system or local runtime state, save the exact command, backup path, and restore procedure immediately. Skip whatever the next session can re-derive by reading the repo. Omitting project stores it in the current directory's workspace, which is right for anything specific to this codebase; pass project='playbook' instead when the lesson holds anywhere, because playbook is searched from every directory while a workspace is not. Set importance to preference for a correction or a stated preference, decision for a chosen approach, knowledge for a stable fact, log for routine detail. Pass supersedes=<id> when this replaces an existing memory instead of adding to it. Credentials, tokens and passwords go to secret_set, never here.
Hybrid memory search. Search before guessing at a port, path, config value, or an earlier decision: a stored answer beats a plausible one. Omit project to use the current workspace; project=all explicitly searches across projects.
Enum parameters lack descriptions of valid values. importance and memory_kind have enums in schema but descriptions do not explain when to use each value (e.g., 'preference' vs 'decision' vs 'knowledge').
Output schemas are not documented. Tools return structured responses but LLMs cannot see what fields to expect (e.g., memory_search returns results but structure is undocumented).
No error recovery guidance. Errors are returned as plain text (e.g., 'memnest unreachable') with no actionable next steps or categorization (retryable vs user-fixable vs fatal).
Destructive operation (memory_delete) lacks confirmation or dry-run support. No pattern to prevent accidental deletion.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Retrieve a secret by exact ID. Returns the plaintext secret value. Secrets are never returned by search or memory_get.
Store a secret (credential, token, password) securely with encryption. Secrets are never returned by search or get; they are only accessible via secret_get with the exact ID.
Parameter constraints not reflected in descriptions. n_results has min/max in schema (1-50) but description does not state the bounds; max_chars has min/max (1-30000) but description omits this.