Cognitive Memory

**Judgement:** High-functionality, opinionated cognitive memory system for Claw agents, suited to advanced, long-lived or multi-agent setups. Very powerful but complex; best for users comfortable with custom file layouts, reflection workflows, and token-budget management. **Strengths** - Deep architecture: four distinct memory stores (episodic, semantic, procedural, vault) plus core MEMORY.md always in context. - Natural-language triggers for remember/forget/reflect with LLM-based classification and targeted writes. - Explicit decay model with clear thresholds and type weights to manage relevance and token economy. - Rich reflection pipeline (token request → internal monologue → approval → consolidation) that feeds IDENTITY.md, SOUL.md, and evolution.md. - Strong auditability: git-based history, audit.log, structured actors, and critical-file alerts. - Multi-agent pattern: shared read, gated write via pending-memories proposals reviewed by a main agent. **Key Risks / Limitations** - **Complexity & setup overhead:** Requires specific directory structure, init scripts, config changes, and adoption of the provided workflows; not a drop-in “just store embeddings” solution. - **Token and latency cost:** Reflection cycles (30K-token inputs, ~8K outputs) and large core memory/context can significantly increase cost and latency if overused. - **LLM routing reliability:** Correct classification of triggers and store selection depends on prompts and model behavior; mis-routed or over-written memories are possible. - **Safety/anthropomorphism concerns:** Files like IDENTITY.md and SOUL.md plus philosophical reflection and “self-image” evolution can encourage users to anthropomorphize the agent or construct brittle identity prompts that increase prompt-injection or manipulation risk. - **Privacy & retention:** Episodic logs, vault entries, git history, and audit logs may accumulate sensitive data; requires careful operational policies for redaction, rotation, and access control. **Recommended Scenarios** - Long-lived assistants or project agents where nuanced, evolving memory and explicit reflection are core features. - Multi-agent workspaces needing shared read/gated write memory with clear provenance and audit trails. - Experimental setups studying reflective behavior, self-modeling, or complex memory strategies. **Less Suitable For** - Lightweight, low-latency bots that only need simple, cheap retrieval. - Environments with strict data-minimization or where maintaining detailed episodic logs and git history is not acceptable without additional privacy controls.