Question 1

What is Dhee?

Accepted Answer

Dhee is a context manager for coding agents. It sits between your agent and the model, decides what context should be admitted, keeps useful memory linked to evidence, and suppresses stale transcript and tool-output bloat.

Question 2

How does Dhee reduce Claude Code token usage?

Accepted Answer

Two mechanisms. First, a 500-line CLAUDE.md that normally re-loads every turn becomes a vector-indexed memory that injects only matching rules — about 240 tokens instead of 5,700. Second, the router wraps Read and Bash so a 10 MB git log becomes a 40-token digest with a pointer; raw content only re-enters context if the model explicitly expands the pointer. Typical reduction is 90%+.

Question 3

How is Dhee different from Mem0, Letta, MemPalace, and agentmemory?

Accepted Answer

Dhee is the only agent memory layer that leads LongMemEval at R@5 99.4% on the full 500-question set, self-evolves its retrieval policy per tool and per intent, and ships a context router that digests Read, Bash, and subagent output at source instead of dumping raw output into context.

Question 4

Does Dhee work with Cursor, Codex, Gemini CLI, Aider, Cline, or Goose?

Accepted Answer

Yes. Dhee exposes an MCP server that plugs into any MCP-compatible client — Claude Code, Cursor, Codex, Gemini CLI, Cline, Goose, Claude Desktop — plus native Claude Code hooks, a Python SDK, a CLI, and Docker. One install, every host.

Question 5

Is Dhee production-ready? What storage does it use?

Accepted Answer

Dhee runs on SQLite by default — no Postgres, no Qdrant, no pgvector, no extra infrastructure. It has 1000+ tests, reproducible benchmarks committed in-tree, an MIT license, and works offline with Ollama embeddings or online with OpenAI, NVIDIA NIM, or Google Gemini.

Question 6

How does Dhee's self-evolution actually work?

Accepted Answer

Every time the model calls dhee_expand_result(ptr) to see the raw output behind a digest, Dhee logs the tool, intent, and depth. The command dhee router tune reads that ledger and atomically rewrites ~/.dhee/router_policy.json — deeper digests for file types the model keeps expanding, shallower ones it never does. No config to hand-maintain.