Bellamente v0.0.1 — Memoria Viva for your AI agents

Bellamente v0.0.1 — Memoria Viva for your AI agents

First release. Bellamente is a local-first memory tool for AI agents: it stores durable facts and
documents, recalls them semantically, and sits in front of your local LLM as a drop-in
/v1/chat/completions proxy that injects relevant memory — with a durable trace of exactly what it
did. One binary. No Docker, no account, no cloud.

Highlights

• Memory you can inspect and trust. Every recall is logged: what was searched, what matched, at
  what score, in how many milliseconds — readable via /inspect and the built-in dashboard.
  Corrections are versioned (nothing silently overwritten), forgetting is reversible and audited.
• Drop-in proxy for any OpenAI-compatible client. Change one base URL. Works with Ollama,
  LM Studio, llama.cpp, vLLM. Memory grounding runs on buffered and streamed requests — the
  proxy intercepts the model's searchMemory call (even mid-stream, even after preamble text),
  runs local recall, re-invokes your model, and streams only the final answer.
• Auto-capture with LLM distillation. Durable first-person facts from your chats are remembered
  through the standard dedup path — extracted by a small LLM pass through the SAME local upstream
  (never a cloud call), with heuristic fallback, a sensitive-content exclusion list
  (credentials/financial/medical are never stored), and a capture trace for every event.
  BELLA_PROXY_CAPTURE=0 disables; BELLA_CAPTURE_DISTILL=0 keeps capture heuristics-only.
• Embedded everything. Postgres + pgvector compiled to WASM run in-process (no server);
  embeddings are local and device-scaled (multilingual-e5-small on capable machines, a static
  Model2Vec model on low-RAM machines — pinned per data dir so hardware changes never corrupt
  recall). Append-only migrations mean upgrades never strand your data.
• Local-first by default. Binds 127.0.0.1 unless you opt out. No telemetry. Your machine is
  the cloud.

Also in this release

• Document ingestion (POST /documents): structure-aware markdown chunking, token-budget guarded;
  hybrid document search (vector + full-text, RRF-fused).
• Full memory lifecycle API: dedup/supersede on write, version-chain reads, PATCH-as-new-version,
  reversible forget, hard delete.
• Upstream timeouts + graceful degradation: a hung model server or failed recall never hangs or
  500s your chat turn.
• Dashboard: traces, search playground, memory browser with edit/forget/delete/history — styled to
  match the website ("La Macchina").

Install

Download the binary for your platform below, chmod +x (macOS/Linux), and run:

bella doctor   # verify: DB, model, ports, disk
bella          # serve on 127.0.0.1:8080

Point your client at http://127.0.0.1:8080/v1. Zero config: no .env, no API key on
loopback (the default). Exposing beyond loopback (BELLA_HOST) auto-generates a key in the data
dir and requires it; BELLA_API_KEY overrides. All other BELLA_* env vars are optional tuning.

Checksums: SHA256SUMS.txt.