consol

Local, inspectable memory engine with compounding knowledge transfer for MCP agents.

Most agent memory frameworks just dump raw transcripts into a vector DB. They burn 5–15k tokens per search, suffer high query latency, and when the agent encounters an unseen task, it fails completely.

Consol is a 100% local, high-performance MCP memory server that enables Compounding Knowledge Transfer — allowing agents to learn like humans by transferring principles across domains.

The Core Innovation: Compounding Knowledge Transfer

Standard RAG searches for direct keyword or embedding overlaps. If you teach an agent to grow 🍌 banana and 🍓 strawberry. When tasked with 🍉 watermelon, standard RAG returns empty or irrelevant matches.

Consol operates on an abstraction hierarchy:

$$\text{Principle} > \text{Pattern} > \text{Specific}$$

1. Lineage Indexing: When an agent solves multiple tasks, Consol distills the shared heuristics (e.g. deep irrigation, root moisture retention, soil balancing).
2. Novelty-Weighted Transfer Boost: When lexical coverage is low (an unseen task like 🍉 watermelon), Consol automatically boosts high-abstraction principles derived from multiple distinct root experiences (🍌 + 🍓).
3. Compounding Capability: Every validated lesson becomes a reusable prior for all future tasks. The agent gets exponentially smarter over months and years with zero retraining.

Why Consol Crushes Everything Else

• ⚡ 10x Faster: ~30ms warm query latency (local Q8 sqlite-vec + FTS5 BM25, zero cloud API calls).

• 💰 5x Token Savings: Bounded candidate descriptors (<1k tokens/recall vs 5–10k in naive RAG).

• 🔍 Zero Black Box: Plain Markdown notes + [[wikilinks]]. Open the vault in Obsidian to see your agent's knowledge graph expand live:

• 🔒 100% Local & Inspectable: Back up to Git, edit notes by hand, zero Docker/Redis/cloud DBs.

• 👥 Multi-Agent Multiplexing: Switch between specialist banks (e.g. coder, researcher, architect) in one MCP server.

• 🛡️ Outcome-Gated Evolution: Inferred experiences start as candidate. Promoted to active only after 2 independent successful runs in production.

30-Second Setup

Paste this prompt to your AI coding agent (Claude Code, Cursor, Codex) — customize the options to your liking:

Read https://raw.githubusercontent.com/keb-org/consol/main/SETUP.md and set up consol memory for this environment:
- Vault: <default: ~/.consol-vault>
- Agents: <default: jarvis>

Your agent will automatically install the package, initialize your local vault, configure multi-agent routing, and wire the MCP server — the server itself injects the mandatory human-like memory protocol (recall → read if confident, search if not) into every connected agent and subagent. No extra host file required.

Host rule is optional hardening: mirroring SETUP.md §5 into CLAUDE.md/AGENTS.md adds per-turn re-injection, but install is one step.

License

MIT © Kryat