Memanto

Memanto is a memory agent: a separate agent that runs beside a fleet of AI agents and manages their long-term memory, deciding what to keep, what conflicts, what expires, and which agent needs to know. It is positioned above storage substrates such as vector stores, filesystems, and platform-native memory features, which it treats as pluggable backends.

Six behaviours each have a command behind them. It observes interaction streams and extracts durable knowledge (decisions, preferences, facts, failures) instead of archiving transcripts (memanto remember --from-conversation); consolidates extracted memories so duplicates collapse and repeated observations raise confidence; reconciles contradictions by superseding rather than appending, so "what is true now" and "what did we believe then" remain separate questions (memanto conflicts); applies decay, expiry, and deletion as policy (memanto forget); briefs each agent with the relevant slice of the estate before it acts (memanto agent bootstrap); and moves knowledge between frameworks and vendors through the Open Knowledge Format (memanto memory export --okf). memanto schedule enable runs curation, merging, and contradiction flagging daily.

memanto connect wires the shared estate into Claude Code, Codex, Cursor, Windsurf, Antigravity, Gemini CLI, Cline, Continue, OpenCode, Goose, Roo, GitHub Copilot, or Augment with no code changes. Memories are typed into 13 categories such as instruction, fact, decision, preference, and error, each agent has its own namespace, and every memory carries a confidence score, source, provenance, and timestamp. recall supports point-in-time queries with --as-of and --changed-since, and answer returns grounded answers from retrieved memory. Files (.pdf, .docx, .xlsx, .json, .txt, .csv, .md) can be ingested into a namespace.

Expiry is governed by per-agent YAML policies with a per-type retention table and named rules where the first match wins; a memory is either active or expired, expired memories still appear in recall labelled [EXPIRED] with the rule that retired them, and memanto memory restore reverses it. The estate exports as plain Markdown in OKF, and memanto migrate imports from Mem0, Letta, Supermemory, or an OKF bundle. Memanto runs fully local with Docker and Ollama, on a free cloud backend, or on your own hosting, switched with memanto config backend. A TypeScript SDK, a REST API (memanto serve), and a local dashboard (memanto ui) are included.

On public recall benchmarks the project reports 89.8% on LongMemEval and 87.1% on LoCoMo, with datasets and harness published on Hugging Face, and notes that cross-project scores are not comparable because reader model, judge model, judge prompt, and retrieval budget each move results.

Features

• Extraction from conversations: durable decisions, preferences, facts, and failures pulled from agent traffic
• Consolidation: duplicates merge and repeated observations strengthen confidence on a daily schedule
• Reconciliation: contradictions supersede rather than append, with both versions kept and surfaced for review
• Policy-driven expiry: per-agent YAML retention tables and named rules; expired memories stay recallable and restorable
• Briefing: agent bootstrap hands each agent the relevant slice before it acts
• Open Knowledge Format: portable Markdown export and import; migrate moves estates from Mem0, Letta, or Supermemory
• Agent connectors: Claude Code, Codex, Cursor, Windsurf, Antigravity, Gemini CLI, Cline, Continue, OpenCode, Goose, Roo, Copilot, and Augment
• Typed memories: 13 categories with per-agent namespaces and per-memory confidence, source, provenance, and timestamp
• Temporal recall: --as-of and --changed-since queries plus grounded answer
• Deployment options: local Docker with Ollama, free cloud, or self-hosting, switchable in one command