| name | llm-wiki |
| description | LLM-wiki is a curated, living knowledge base about LLM systems programming, prompt engineering, and long-term memory strategies. Registered as a reference skill to lift OctoCode's systems-programming and memory design. |
| source | https://github.com/rasbt/LLMs-from-scratch |
| integration | external-knowledge |
| status | descriptor-only |
llm-wiki
Descriptor-only registration (P12 batch). Used as a reference body of knowledge for:
- Systems programming uplift — tokenization, attention, KV-cache, batching, quantization, and how each interacts with coordinator latency budgets.
- Memory uplift — session store layering (OctoCode's
FileSessionStore + SQLite persistence), context-window budgeting, and semantic vs. episodic memory separation.
Why registered
OctoCode's runtime already separates SessionStore, ConversationStore, and TurnStateStore; the llm-wiki material is a rigorous reference for extending those stores with:
- Episodic memory — per-session event journal (already partially present via
event_feed_json).
- Semantic memory — embeddings-backed retrieval layered above the P9 token index (
crates/octocode-cli/src/index.rs).
- Procedural memory — skill index itself (this directory).
Integration plan (deferred to P13+)
- Add a
crates/octocode-runtime/src/memory.rs module that unifies the three stores behind a MemoryLayer trait.
- Expose
octocode-cli memory-stats subcommand emitting counts and byte sizes per layer.
- Wire semantic retrieval through a pluggable embedding provider (reuses
ProviderFactory).
Safety notes
Embedding calls are outbound network traffic — gate by the same allowlist governance as skills-install.
