Research exchange: 5 cognitive memory papers + OpenClaw fork implementation

[globalcaos]

Hey MemOS team 👋

We noticed your arXiv paper (2507.03724) and the impressive benchmarks (+43.7% vs OpenAI Memory, 35% token savings). We've been pursuing similar research from a different angle and thought a comparison could benefit both projects.

Our Work

We maintain an OpenClaw fork with a cognitive memory architecture (~150 files, 7 modules). We've written 5 academic papers, each covering a different aspect of agent memory:

Paper	Module	Key Contribution
ENGRAM	Context compaction	Treats context limits as cache eviction (LRU/LFU-inspired), not summarization. Sleep consolidation pipeline.
CORTEX	Agent identity	EWMA SyncScore measures persona drift mid-conversation. Auto re-injection when drift > threshold.
HIPPOCAMPUS	Retrieval indexing	Pre-computed concept index (500 anchors → 9500 chunks). O(1) lookup vs runtime vector search.
LIMBIC	Humor detection	Bisociation detection in embedding space (Koestler's theory formalized).
SYNAPSE	Multi-model reasoning	RAAC debate protocol with cognitive diversity scoring.

Overlap with MemOS

Your memory type taxonomy (working, episodic, semantic, procedural, strategic) is nearly identical to our ENGRAM's — we'd love to compare the implementation approaches:

• Your MemOS Cloud API vs our SQLite-native zero-infra approach — different trade-offs worth benchmarking
• Your token savings claims vs our HIPPOCAMPUS pre-computed indexing — both reduce inference-time cost, different mechanisms
• Skill memory (your unique contribution) vs our procedural memory — how do they compare on cross-task reuse?

What We're Exploring Now

Context Anatomy — per-turn visual decomposition of the prompt: system prompt vs. history vs. tools vs. retrieval. A debugger for what the model actually sees.

Curiosity-Driven Exploration — agent autonomously identifies knowledge gaps and generates questions. Proactive memory, not just reactive.

Proposal

1. Benchmark comparison on LoCoMo / LongMemEval using both systems
2. Architecture deep-dive — we share our papers, you share insights from your arXiv work
3. Joint exploration of pre-computed indexing + your token optimization

Fork: globalcaos/tinkerclaw

Happy to share full PDFs of all 5 papers. 🤝

[CaralHsi]

Hi @globalcaos — thanks for sharing the papers and the detailed proposal.

The ENGRAM/CORTEX/HIPPOCAMPUS/LIMBIC/SYNAPSE breakdown and the overlap with our memory taxonomy are genuinely interesting, especially given the different implementation trade-offs (Cloud API vs SQLite-native, pre-computed indexing, procedural/skill memory).

We’ll keep following your work and appreciate you putting this together — feel free to keep us posted as you publish updates or new benchmark results.

[globalcaos]

Thanks for the thoughtful response, @CaralHsi — and apologies for the delayed follow-up.

Quick update: Our fork has been renamed and restructured since the original post:

🔗 New repo: globalcaos/tinkerclaw

The cognitive memory modules (ENGRAM, CORTEX, HIPPOCAMPUS, LIMBIC, SYNAPSE) are all active and have been running in production for 8+ weeks now. A few concrete results since we last posted:

• ENGRAM compaction reduced average context window from 23.5KB to 12KB with no measurable quality loss in long conversations
• HIPPOCAMPUS indexing now covers 9,500+ chunks across 500 concept anchors — retrieval is sub-millisecond
• Context Anatomy (the visualization tool) is live in our Tinker UI, decomposing every prompt into system/history/tools/retrieval segments with real-time cost tracking

We'd still love to run a benchmark comparison on LoCoMo or LongMemEval if there's interest on your end. We can set up a standardized evaluation harness and share results publicly.

Also published a Field Guide — 32 operational lessons from running the system 24/7 with the full memory stack. Might be useful reference for MemOS users running long-lived agents.

Happy to share the full paper PDFs anytime. 🤝

---

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

[globalcaos]

Hi @CaralHsi — following up on the benchmark proposal from our last message. We've been running the LoCoMo evaluation setup locally and could share the harness config. Is there a preferred timeline or format for the comparison on your end? Happy to adapt to whatever works for the MemOS team.

🔗 Fork (tinkerclaw): https://github.com/globalcaos/tinkerclaw

[github-actions[bot]]

This issue has been automatically marked as stale due to inactivity.

[CarltonXiang]

Thanks for the continued discussion here.

Since this topic is better suited for research and benchmark follow-up than a scoped engineering issue, we have moved the conversation to GitHub Discussions.

Please continue sharing the benchmark scope, datasets, metrics, comparison targets, and any updates in the Discussion thread:

If the discussion later converges into a concrete MemOS implementation task with clear acceptance criteria, we can open a dedicated issue from there.

Closing this issue to keep the issue tracker focused on actionable bugs and feature work.