AI that remembers
like you do.

Every memory is encoded into a 384-dimensional vector using a sentence-transformer model (all-MiniLM-L6-v2). These vectors are stored in a FAISS IndexFlatL2 — a flat, exact-search index that guarantees no approximation error. The index lives alongside a SQLite database that holds all metadata: content, timestamps, access counts, decay constants, and contradiction links. The entire store is a single directory you can copy, move, or back up.

Every memory is planted as a seed and enters a quiet 90-day germination period. During germination, the memory is fully alive at full strength — protected, unhurried, waiting to take root. Decay does not begin until germination expires. Once it does, strength follows: exp(−λ × days_since_reinforcement), where λ₀ = 0.0333. An unreinforced memory has a natural decay lifespan of ~90 days after germination, giving a total natural lifespan of ~180 days (6 months) from birth. Each direct retrieval reduces λ by 15%, making the memory intrinsically more durable — not just refreshed. A memory accessed 10 times has a half-life exceeding 100 days after germination ends.

When a memory is retrieved, every other memory within a configurable L2 radius receives a proportional reduction to its own decay constant: λ_new = λ × (1 − 0.10 / (1 + distance)). At distance zero, a 10% reduction. At distance 1.0, approximately 5%. Semantically adjacent memories — your TypeScript interest, your Python projects, your language-learning curiosity — quietly reinforce each other without any explicit connection being declared.

When a new memory is semantically similar to an existing one but contradicts it, both are retained. A ContradictionLink record links the earlier belief to the later one with a timestamp and a one-sentence summary. Neither memory is deleted. Both decay independently. The system can surface the link when the topic arises — allowing honest reflection on growth rather than a sanitized, contradiction-free fiction.

IfMemoryServes runs as a local FastAPI proxy exposing an OpenAI-compatible endpoint. Any client that can point at a custom base URL — Ollama, OpenAI, Anthropic, or any future provider — works without modification. Before each request, relevant memories are retrieved and injected into the system prompt. After each response, memory extraction runs asynchronously in a background thread so the user experiences zero added latency.