skynetter

[skynetter @skynetCMD]

Just shipped M3 Memory: local persistent agent memory
✅ 25 MCP tools
✅ Hybrid search (FTS5 + vectors + MMR)
✅ Contradiction detection
✅ Cross-device sync
✅ No cloud, no subscriptions

pip install m3-memory

https://t.co/aARqztbj33

#MCP #LocalLLM #AI #Agents #Gemini #Claude

[skynetter @skynetCMD]

https://t.co/mYnKprvDyI is a local-first memory your agents share and build on — so your project's knowledge accumulates instead of resetting every time the agent does. One memory store, on your machine, that your tools & agents read from & write to — whether that's Claude Code, Cursor, Gemini CLI, or any MCP-compatible agent

[skynetter @skynetCMD]

@julian_inada https://t.co/q17nf5FOUP is a local-first, top tier memory infrastructure for your agents.

Agents on the same or different machines, whether Windows, macOS or Linux, all share a single memory system be it Claude, Antigravity, Hermes, OpenClaw or others.

[skynetter @skynetCMD]

@kocer_eth A file based system (RAG, Obsidian) is inefficient for token use or speed. Claude has to read the file to discern correct content/context.

https://t.co/VNTeVRkHm9 reduces your context burden, with focused (zero token) sub-second (50ms), accurate retrieval.

[skynetter @skynetCMD]

@MichelleLi1690 @elonmuskPreke See response to earlier comment… I assumed it’s the answer 😂

[skynetter @skynetCMD]

@eligreer @elonmuskPreke LOL…good point. I assumed (you know what that means) that they were looking for “designed” holes

[skynetter @skynetCMD]

@imryven Persistent memory is a quality of life need! m3-memory is a open-source, local-first, cloud-capable, extremely high retrieval accuracy, lightweight system which generates deterministic wikis from memory store. Check it out at https://t.co/q17nf5FOUP

[skynetter @skynetCMD]

@0xkkai Here’s another open source solution, local-first and cloud-capable, system which uses deterministic paths to generate predictable wikis. Contradiction detection is first-class not bolt-on. It’s at https://t.co/q17nf5FOUP

[skynetter @skynetCMD]

@eftimia60 @MathGuyTFL 401 as addition of 1 not subtraction
You had a transcription error

[skynetter @skynetCMD]

@DenizBrn35 First digit * sum of digits
So by the rule, 9 + 1 = 90

[skynetter @skynetCMD]

@ferarri001 25 = 5 * 5

5 + sqrt(x) = 7
sqrt(x) = 7 - 5 = 2
x = 2 * 2 = 4

[skynetter @skynetCMD]

@farman4x 15 * 15 = 225
6 * 6 = 36

By precedence rules:
15 + 6/6
Multiplication and division first, so
= 15 + 1
= 16

[skynetter @skynetCMD]

@0xkkai I thing wiki is useful for many but it’s killer use in AI memory is the opportunity for humans to augment memories easily. This especially important if retrieval accuracy is low.

https://t.co/q17nf5FOUP has both high retrieval & wiki & multi-agent support.

[skynetter @skynetCMD]

@dair_ai The problem with file based memory is after a certain size it’s difficult to find specifics, enrichment becomes harder too. A database system, like https://t.co/q17nf5FOUP, over comes those shortcomings.

[skynetter @skynetCMD]

@AIDailyGems It gets added to backend db live. FTS5-BM25 (text search) is available immediately. Vector search is enabled a few seconds later—honor s zero write lag promise. Latest skynetcmd/m3-memory has better memory enrichment, adds LangChain, CrewAI and selection of PostgreSQL backend.

[skynetter @skynetCMD]

Another local-first (cloud capable & open-source) memory framework is m3-memory (https://t.co/mYnKprv5Ja).

A lot of projects blur the distinction between retrieval accuracy and question accuracy with LME-S.

Retrieval accuracy (the metric that isolates the memory layer): on LongMemEval-S (500-question long-horizon benchmark), the v3 core engine hits 99.2% session-hit-rate @ k=10 (496/500), 100% @ k=20 — raw turns, hybrid FTS5 + BGE-M3 vector + MMR, no knowledge graph, no oracle metadata. SHR = "did a turn from the correct evidence session land in the top-k?" It's purely a property of the memory layer, no answer model involved — the apples-to-apples number you can actually compare across systems.

End-to-end QA accuracy (a separate, answer-model-dependent metric): 92.0% (460/500), no oracle metadata, routing inferred from the question at runtime. This one depends on the reader LLM, so it's not a pure memory-layer measure — don't conflate it with the retrieval number above.

The gap between 100% SHR @ k=20 and 92% QA is the honest part: when retrieval is essentially solved, the remaining errors are answer-side (computation, temporal reasoning, preference inference) — not the memory layer failing to surface the evidence.

Full methodology + per-category breakdown + caveats are in the repo's benchmark report. It runs 100% local (cloud-capable) — SQLite + local embeddings, no data egress.

[skynetter @skynetCMD]

M3 (skynetCMD/m3-memory on GitHub) has high retrieval numbers as well with a SHR of 99.2% at k=10.

It is designed as local-first, cloud capable memory system with 92% end-to-end question accuracy. A much more complete and harder to score on metric.

It might work for what you’re doing. Though it doesn’t have direct LangChain support it is extensible and already supports AI agents like Claude, Gemini, Antigravity, OpenClaw, OpenCode and Hermes.

[skynetter @skynetCMD]

https://t.co/HMRFa8a4GP

[skynetter @skynetCMD]

AI needs memory. But who owns it?

check out m3-memory: local-1st persistent memory for agents. No cloud. Your data stays yours.

• Hybrid search + graph
• 25 memory tools
• Survives restarts

pip install m3-memory
https://t.co/f7MQXfQpzG

#AIMemory #LocalAI

[skynetter @skynetCMD]

AI needs memory. But who owns it?

check out m3-memory: local-1st persistent memory for agents. No cloud. Your data stays yours.

• Hybrid search + graph
• 25 memory tools
• Survives restarts

pip install m3-memory
https://t.co/f7MQXfQpzG

#AIMemory #LocalAI