medyll
Profile
← all projects

humemory

Shared agent memory with decay, cues and reviewed routines

Started Last update

humemory gives Claude, Codex, Kimi Code and OpenCode a shared SQLite memory that behaves less like a log and more like recall. Decaying traces, lexical and vector retrieval, prospective cues and reviewed cognitive scripts work offline; JEV can shadow-qualify candidate routines, while optional model-assisted maintenance runs outside the agent path.

Where it sits


flowchart TD
  P["idea / needs"] --> S["humemory"]
  S --> A["retrospective · decay + lexical/vector retrieval"]
  S --> B["prospective · cues + open loops"]
  S --> C["trusted shared memory · MCP + provenance"]
  S --> D["off-path maintenance · deterministic fallback"]
  D --> E["dreaming + reviewed cognitive scripts"]

Overview


Memory engine for AI agents modeled on human recall. It keeps decaying traces, open intentions, provenance and reviewed cognitive scripts in SQLite, then retrieves them through lexical and vector search. Claude hooks plus explicit Codex, Kimi Code and OpenCode importers feed one MCP store; background maintenance stays off the agent path and falls back to deterministic local processing when no model is available.

Highlights


  • · Offline core with optional model-assisted maintenance
  • · Shared MCP store · provenance · contradiction handling
  • · JEV shadow qualification for reviewed cognitive routines
TypeScriptBunSQLiteMCPONNXReact
View source →