1. Self-Evolving Agent Frameworks (skills + memory + evolution loop): MetaClaw (proxy + RL weight updates), OpenSpace (MCP skill engine), Hermes (full agent with curated memory)
2. Standalone Memory Layers (bolt-on to any agent): Mem0 (vector + graph, 48k★), Zep/Graphiti (temporal knowledge graph), SimpleMem (semantic compression, +64% over Claude-Mem)
3. Agent Runtime with Memory OS: Letta (MemGPT — self-editing memory, LLM-as-OS)
For swarm.ing: Combine OpenSpace (skill evolution via MCP) + Mem0 or SimpleMem (factual memory layer) + steal Letta's self-editing pattern for your ContextManager.
| Dimension | MetaClaw 🦞 | OpenSpace 🧬 | Hermes 🪶 | Mem0 🧠 | Letta 🖥️ | Zep ⏱️ | SimpleMem 📦 |
|---|---|---|---|---|---|---|---|
| Identity | |||||||
| Origin | AIMING Lab, UNC Chapel Hill | HKUDS, Univ. of Hong Kong | Nous Research | Mem0.ai (YC, $24M raised) | UC Berkeley → Letta Inc | Zep AI (Graphiti engine) | AIMING Lab, UNC |
| Type | Proxy + RL meta-learner | MCP skill engine | Full agent framework | Standalone memory layer | Agent runtime (LLM OS) | Temporal memory service | Memory compression lib |
| GitHub ★ | ~3.2k | ~5k+ | ~8.8k | ~48k TOP | ~38k | ~4k | ~3.2k |
| License | MIT | MIT | MIT | Apache 2.0 | Apache 2.0 | Apache 2.0 | MIT |
| Memory Architecture | |||||||
| Memory Model | SkillBank (behavioral) + Contexture (factual, v0.4). Dual: skills + facts | SQLite SkillStore with lineage DAG. Skills ARE the memory | Multi-layer: curated facts + FTS5 session index + Honcho user model + skills | Tri-store: vector + knowledge graph + KV. User/session/agent scopes WIN | 3-tier OS: Core (in-context, editable) → Recall → Archival WIN | Temporal KG via Graphiti. Nodes=entities, edges with time validity WIN | 3-stage: Compression → Synthesis → Intent-Aware Retrieval |
| Self-Editing? | Auto via LLM evolver + background consolidation | Auto FIX/DERIVED/CAPTURED | Agent adds, replaces, consolidates. Security scanning | Adaptive — corrects existing memories | Core innovation — agent edits own context blocks WIN | Auto entity extraction + temporal edges | Online synthesis merges fragments at write time |
| Temporal | Skill versioning | Lineage DAG + timestamps | Session timestamps + FTS5 | Creation timestamps only | Edit timestamps. No temporal reasoning | First-class — valid_from/to/invalid_at WIN | Absolute timestamps on units |
| Knowledge Graph | No | Skill lineage graph | No (file-based) | Yes — Neo4j/Memgraph WIN | No (structured blocks) | Core IS a temporal KG WIN | No (triple-indexed units) |
| Retrieval & Efficiency | |||||||
| Retrieval | Top-k skill retrieval | BM25 + embedding → LLM select | FTS5 + LLM summarization | Vector + graph + reranking | Agent-driven tool calls | Graph walk + semantic triplets | Intent-aware parallel multi-view WIN |
| Token Efficiency | 10-20% compression | 46% savings (GDPVal) WIN | Progressive L0→L1→L2 | 90% fewer vs full-context | Virtual context paging | Episode summaries | ~550 tokens for 43.24% F1 WIN |
| Skill Evolution | |||||||
| Evolution | Dual-loop: Fast (skill synth) + Slow (RL LoRA) WIN | FIX/DERIVED/CAPTURED × 3 triggers | Self-improving loop every 15 tool calls | Memory updates only | Letta Code: skills + subagents + continual learning | Graph evolves with episodes | No skill evolution |
| Weight Updates? | Yes — cloud LoRA via Tinker WIN | No | No (Atropos offline) | No | No | No | No |
| Community | openclaw-skills | open-space.cloud WIN | agentskills.io + Skills Hub | N/A | Emerging | N/A | N/A |
| Integration | |||||||
| Integration | Proxy wrapping OpenClaw agents | MCP server → any SKILL.md agent WIN | Standalone + OpenClaw import | SDK bolt-on to ANY framework WIN | Full runtime — agents inside Letta | SDK/API bolt-on | pip + MCP server |
| CF Workers Fit | Persistent proxy needed | MCP→ToolRegistry. SQLite→D1 WIN | External via Tailscale | API calls from DO→Mem0 | External — heavy runtime | API from DO→Zep Cloud | Embed in DO possible |
| Benchmarks | |||||||
| Key Result | K2.5: 21→40.6%. +185% file-check | 4.2× income. 70.8% quality. -46% tokens | 8.8k★, 142 contributors | +26% over OpenAI Mem. -90% tokens | #1 on Terminal-Bench | 63.8% LongMemEval WIN | +64% over Claude-Mem WIN |
Skill Layer: OpenSpace via MCP → ToolRegistry. FIX/DERIVED/CAPTURED maps to SKILL.md. Skill DB → D1.
Factual Memory: SimpleMem for per-user context (~550 tokens). Or Mem0 if you need graph relationships.
Temporal Layer: Zep/Graphiti for trading temporal reasoning — position history, stop-loss changes, fact validity windows.
Patterns to steal: Letta's self-editing memory → ContextManager. Hermes's L0→L1→L2 progressive disclosure. MetaClaw's OMLS idle scheduler for future LoRA.
All 7 validate: harness engineering > model selection. The moat is in the orchestration.