aayoawoyemi/Ori-Mnemos

Local-first persistent agentic memory powered by Recursive Memory Harness (RMH). Open source must win.

318 ★27 forksTypeScriptUpdated 7d ago

What you need to know

Ori Mnemos — open-source persistent memory infrastructure for AI agents implemented as cognitive models on a knowledge graph (markdown files + wiki-links + SQLite index): ACT-R activation decay, spreading activation, Hebbian co-occurrence, and RL on retrieval. Zero cloud dependencies, zero API keys required for core functionality.

Install

npm install -g ori-memory
ori init my-agent
ori bridge claude-code --vault ~/brain   # full lifecycle: hooks + MCP + CLAUDE.md

Usage

  • Bridge adapters: ori bridge hermes / opencode / cursor / codex / generic --vault <path>
  • Manual MCP config: { mcpServers: { ori: { command: "ori", args: ["serve", "--mcp", "--vault", "/path/to/brain"] } } }
  • ori add <title> to capture to inbox, ori promote to classify and link, ori explore <query> for recursive graph traversal, ori query ranked <query> for full retrieval
  • ori prune --apply to archive stale notes based on activation topology

Key features

  • 16 MCP tools + 5 resources (ori_orient session briefing, ori_query_ranked, ori_explore, ori_prune, ori_index_build, etc.); 17 CLI commands; 579 tests
  • Knowledge graph: wiki-links as directed edges, PageRank authority, Louvain community detection, betweenness, bridge/orphan/dangling analysis
  • Three memory spaces with distinct decay: identity (0.1x), knowledge (1.0x), operations (3.0x); cognitive forgetting via ACT-R base-level equations, Tarjan-protected critical nodes
  • Four-signal fusion retrieval (semantic embeddings + BM25 + personalized PageRank + associative warmth → RRF) with dampening pipeline; Q-value reranking, co-occurrence (Hebbian) edges, LinUCB stage meta-learning
  • Citability: cites 63 research sources; benchmarks claim 90% Recall@5 on HotpotQA (vs Mem0 29%) and 9.5x faster, on plain markdown + SQLite; 'Markdown on disk, Git as version control' sovereignty model

Best for

Agents that need persistent, sovereign, cross-session memory stored as portable markdown/git files — for Claude Code, Hermes, OpenCode, Cursor, or Codex.

Caveats

  • LLM integration is optional — every operation works deterministically with heuristics alone; when configured, LLM improves classification and link suggestions
  • Benchmark latency/quality claims are from the project's own evaluation (GPT-4.1-mini for answer generation)
  • Remote/headless: ori must be on PATH inside the environment and vault on persistent storage (not ephemeral) for serverless backends like Modal
Platforms: macOS · Linux · WindowsClients: Claude Code · Hermes Agent · OpenCode · Cursor · Codex
Documentation ↗

Reviewed 2026-08-07

Topics

agent-memoryai-agentai-agentsai-memoryknowledge-graphllmlocal-firstmarkdownmcpmcp-servermemorymodel-context-protocolpersistent-ai-memorypersistent-memoryrecursive-memory-harnessrmhsemantic-searchsqlitetypescriptvector-search
View on GitHub ↗
Stars
318★
Forks
27
Language
TypeScript
License
Apache-2.0
Created
2026-02-20
Last push
2026-07-30