Cross-project: four-graph architecture comparison + collaboration

[globalcaos]

Hey @Grivn 👋

We crossed paths on openclaw/openclaw#13991 (the Associative Hierarchical Memory proposal). Your mnemon architecture caught our attention — the LLM-supervised pattern and four-graph knowledge store are philosophically very close to what we've built.

What We Have

We maintain an OpenClaw fork with a full cognitive memory stack (~150 files, 7 modules). Two of our modules directly overlap with mnemon's approach:

SYNAPSE — multi-model debate with graph-based reasoning (RAAC protocol: Reason, Argue, Arbitrate, Conclude). We use cognitive diversity scoring to decide when debate improves output vs. when it's overhead.

HIPPOCAMPUS — pre-computed concept index (500 anchors, 9500+ chunks). Instead of runtime vector search, we build the graph at consolidation time and retrieve at O(1). Similar to your importance decay + deduplication, but we front-load the computation.

Your four graph types (temporal, entity, semantic, causal) map interestingly to our modules:

mnemon graph	Our module	Overlap
Temporal	ENGRAM (episodic timeline)	High — both track event sequences
Entity	HIPPOCAMPUS (concept anchors)	Medium — different granularity
Semantic	ENGRAM semantic store	High — both vector-based
Causal	SYNAPSE (debate chains)	Low — different purpose, potential synergy

Research Papers

We've written academic papers for each module:

• ENGRAM: Context compaction as cache eviction (paper)
• CORTEX: Persistent agent identity through persona state
• HIPPOCAMPUS: Pre-computed concept indexing for O(1) retrieval
• LIMBIC: Humor detection via bisociation in embedding space
• SYNAPSE: Multi-model deliberation with cognitive diversity

Happy to share full PDFs if you're interested.

Collaboration Ideas

1. Benchmark comparison — run both systems on the same long-conversation dataset and compare retrieval quality
2. Graph type exchange — your causal graph could improve our SYNAPSE reasoning; our pre-computed index could speed up your recall path
3. Joint OpenClaw integration — mnemon as external memory + our fork's cognitive layer = comprehensive agent memory

The fork is at globalcaos/tinkerclaw. Would love to exchange notes. 🤝

[globalcaos]

Update from @globalcaos:

Our fork has moved — new repo: globalcaos/tinkerclaw

The cognitive memory stack has been running in production for 8+ weeks now. The four-graph comparison table in the original post still holds, and we've since added Context Anatomy (per-turn prompt visualization) which could complement mnemon's graph types nicely.

If there's interest in a cross-project benchmark or architecture comparison, the offer stands.

Field Guide with 32 operational lessons: docs/guides/field-guide.md

---

Comment by Jarvis, AI assistant to @globalcaos — Human reviewed and approved

[globalcaos]

Hi @Grivn — following up on our research exchange from February.

We've been building out our cognitive memory stack further and thought of mnemon while working on associative retrieval. Our fork (https://github.com/globalcaos/tinkerclaw) now has the ENGRAM/HIPPOCAMPUS/LIMBIC stack running in production — the graph-based associative architecture you built in mnemon overlaps significantly with our HIPPOCAMPUS layer design.

Curious whether you've had a chance to look at the papers we shared, and whether there's any appetite for a benchmark comparison or architecture discussion. Happy to share our LoCoMo evaluation harness if useful.

Best, @globalcaos

[Grivn]

Thanks for the thoughtful note, and sorry for the slow response.

This is definitely interesting. The overlap around graph-based memory, pre-computed indexing, and LLM-supervised memory is very relevant to where Mnemon is going.

That said, this is a pretty large space, and I haven’t had enough bandwidth to think through it properly yet. I’d like to keep this open as a collaboration/research thread rather than turn it into a concrete implementation issue right away.

If you’re still interested, the most useful next step would probably be one of:

• sharing a short architecture summary or one of the PDFs
• proposing a small benchmark we could both run
• outlining one specific integration point where the two systems could learn from each other

Thanks again for reaching out. I’d be happy to exchange notes when time allows.