HyperMnesia

HyperMnesia

*The opposite of amnesia* (Greek *hypermnesia* -- abnormally complete recall). Self-hosted long-term memory for AI coding agents -- a Postgres-backed store that gives an agent (Claude Code, or any MCP client) two things:

1. Architectural memory (doc-RAG + Tier 0/1) -- your repos' docs made searchable, plus a deterministic component -> constraint model so the agent knows the rules that apply to a file before it edits, without a search.
2. Personal memory -- durable facts/preferences/decisions distilled from work sessions and injected back automatically, so the agent stops re-learning the same things every session.

One store (Postgres + [pgvector](https://github.com/pgvector/pgvector)), local-model friendly (embeddings via [Ollama](https://ollama.com) or [TEI](https://github.com/huggingface/text-embeddings-inference)), MCP-native, no cloud dependency. Runs on a laptop, one server, or Kubernetes.

Not a vector-DB wrapper. The value is the **delivery**: deterministic constraint injection (Tier 1) and hook-driven personal-memory capture/recall -- "storage is solved, injection isn't."

How it works

                 +---------------- retrieval ----------------+
   query --> Tier 0: pinned project map + global invariants   |
             Tier 1: path -> component -> must/should constraints (+1-hop graph)   <- deterministic
             Tier 2: hybrid RRF search (pgvector cosine + Postgres FTS) + rerank  <- fuzzy
                 +-------------------------------------------+

• Tier 2 search fuses dense (bge-m3 embeddings, HNSW) and lexical (composite tsvector, works for code identifiers and non-English) via Reciprocal Rank Fusion, then an optional cross-encoder reranker (bge-reranker-v2-m3) reorders the top candidates (measured +0.16 recall@1 on our eval set).
• Personal memory is bi-temporal (event time vs ingestion time), **supersede-not-overwr