A framework-agnostic, fully-offline AI memory system (Python library `taosmd` + optional MCP server) built around 'provable memory': everything lands first in an append-only verbatim archive that is never edited or deleted, and the searchable memory is derived from that archive, never written over it. Because the source is retained, a verifier checks each extracted fact against the exact text it came from and leaves out what it can't support. Part of the taOS ecosystem; runs on 8GB+ RAM (Raspberry Pi 4B to workstation), zero cloud.
- Storage
- Three local SQLite-backed stores: an append-only Archive (verbatim conversation turns, gzipped daily, with an FTS index), a KnowledgeGraph of structured triples with temporal validity (facts are updated, not deleted), and a VectorMemory using all-MiniLM-L6-v2 ONNX embeddings. Per-agent isolation is enforced by an `agent` tag on every row within one shared process.
- Retrieval
- Hybrid keyword + vector search over VectorMemory with pluggable fusion (mem0_additive on CPU; MaxSim + bge-v2-m3 cross-encoder reranking on a GPU), plus archive FTS and KG entity queries. A 'Librarian' layer adds LLM-assisted query expansion to bridge vocabulary gaps. Fact extraction + generation run on a local Qwen3-4B via Ollama (or NPU-optimized RKLLM on RK3588).
- Self-host
- Self-host: moderate
- License
- MIT
- Pricing
- Open-source MIT Python package (`pip install taosmd`), free and fully offline; setup pulls a ~90MB ONNX embedder and a local Qwen3-4B (~2.6GB) via Ollama. No paid tier. · Free / OSS
- GitHub stars
- 60
- Last release
- 2026-06-22
- Last commit
- 2026-06-24
- First catalogued
- 2026-06-28
Strengths
- Zero-loss append-only verbatim archive: the source is never overwritten, so summaries/facts are derived from — and verifiable against — ground truth
- Fact verifier reports the share of extracted facts it cannot support and keeps them out of recall (a measurable unsupported-fact rate, ~18.8% on their data)
- Genuinely low-end and air-gapped: runs offline on 8GB+ RAM (Pi 4B / Orange Pi 5 Plus NPU / mini PC) with no cloud or API keys
Watch out
- README headlines self-reported benchmarks (97.0% Recall@5 on LongMemEval-S; LoCoMo numbers) — record via harvest-benchmarks with selfReported:true, not here. (Notably, the README transparently discloses and corrects an earlier inflated end-to-end Judge figure.)
- GitHub classifies the license 'NOASSERTION', but the LICENSE file is verbatim MIT (Copyright 2026 jaylfc, confirmed)
- Small/young project (~60 stars, created 2026-04); setup expects Ollama + local model downloads and is still being validated across clean machines
Best for
- Offline / air-gapped or low-resource deployments needing auditable, source-preserving memory with no cloud dependency
Benchmark results
No sourced results yet.
Sources
- taOSmd README (vendor)
- GitHub API repo metadata (stars, v0.4.0 release) (third-party)
- LICENSE (verbatim MIT text, Copyright 2026 jaylfc; GitHub auto-classifies NOASSERTION) (vendor)
Last verified 2026-06-28 · updated by discover-frameworks