Memory Mind Mesh — Living Memory for AI

Status: in development, no public artifact yet. The source repo is private and there is no live demo a visitor can interact with. The core feedback loop is now proven by a real, reproduced eval (see “Measured Result” below), but the system stays private until the eval is widened and the public-artifact criteria below are met — so the architecture description reflects a locally-verified, not yet externally-verifiable, state.

A memory system where stored memories learn and evolve through use. Static vector databases return the same thing forever; Memory Mind Mesh uses DSPy modules to refine memory quality over time, driven by feedback from real queries.

What Makes It Different

Living Memories

Memories improve through DSPy optimization keyed on user feedback — responses get more accurate and more succinct with exposure, instead of decaying as the model drifts away from them.

Hybrid Store

Reliable static memory (vector retrieval — cosine over precomputed embeddings) sits alongside living memory (DSPy-optimized). The hybrid router chooses per query, then returns a response with provenance.

Sharded Architecture

Eight shards with ~2000 memories each, distributed with intelligent load balancing. Multi-tier caching and batch processing keep the hot path fast.

Feedback-Driven Evolution

A /memory/feedback endpoint captures gold-standard corrections (query → original response → corrected response → confidence). A scheduled evolution run consumes feedback and recompiles affected memory modules — resource-aware, with automatic throttling to avoid swamping the host.

Quality Assurance

6-dimensional feedback validation catches low-quality or adversarial feedback before it’s used to evolve memories. Consensus management prevents a single contributor from moving a memory’s behavior unilaterally.

Measured Result

The central claim — corrections make recall more accurate and more concise — now passes a real, reproduced eval (eval/receipt.json, status: "pass"). On 10 held-out single-fact memories, after corrections + a DSPy recompile, accuracy (embedding cosine-to-gold) went 0.60 → 0.99 and answer length 12.5 → 3.5 tokens, with 10/10 topics improved and none regressed (LM gpt-4o-mini, temperature 0).

Honest caveats: this is a small, clean test, and reaching it required fixing four independent bugs that had each silently disabled the feedback→evolve loop. Storage is SQLite by default with zero external services; the Postgres + Redis stack is an optional, unverified scale profile. The result is real but private — there is no public artifact to reproduce it against yet.

Technical Stack

• Language: Python
• Orchestration: DSPy (for memory modules + evolution), Ollama (recommended) or OpenAI (optional)
• Indexing: flat cosine over precomputed embeddings (HNSW configured for the scale profile, not yet exercised)
• API: /memory/query, /memory/feedback, /memory/evolve (FastAPI)
• Storage: SQLite by default (zero external services); optional Postgres + Redis + nginx scale profile (unverified)
• Ops: Docker Compose, start_production_system.sh one-command bring-up, production-readiness test suite

Related explorations

Memory Mind Mesh is the flagship of a broader investigation into persistent memory — Aether Memory Stack (tiered cache-style memory), InfiniteContext (progressive summarization for windowless conversations), and Golden Archive (validation-gated knowledge store) are early explorations in the same line. Memory Mind Mesh is the one currently developed here; of the three explorations, InfiniteContext is public (source on GitHub), while Aether Memory Stack and Golden Archive are not publicly distributed.

What’s Needed For This Entry To Tighten

• A public source repository linked via githubUrl, and/or
• A downloadable, usable public artifact (release binary, package, or runnable distribution).