World memory for AI agents.
A arc based context compiler for ai agents
Dhee gives an agent the bigger story before it takes the next action.
LLMs are powerful, but they still miss the arc. What is the user trying to become? What happened last session? Which decision was already made? Which failure should not be repeated? Which proof is required before touching code?
Dhee stores that story locally and compiles it into the smallest useful context. Not a transcript pile. Not vibes. Memory with shape.
pip install dhee
dhee installOr use the one-command installer:
curl -fsSL https://raw.githubusercontent.com/Sankhya-AI/Dhee/main/install.sh | shWire a repo:
cd /path/to/repo
dhee init
dhee status
dhee uiDhee is the memory layer agents should have had from day one.
It runs beside the model and answers the questions that raw chat history cannot:
| The agent needs | Dhee gives it |
|---|---|
| What matters right now? | A compiled context packet with included and rejected reasons. |
| What happened before? | Durable user, project, repo, scene, and handoff memory. |
| What files are risky? | Repo brain impact analysis, likely tests, owners, routes, and symbols. |
| What should not be used? | Secret filters, privacy classes, supersession, contradiction, and proof gates. |
| What is the larger story? | Series, Seasons, Episodes, Scenes, and SceneCards. |
| Can another agent continue? | Session digests, shared task state, evidence refs, and handoff packets. |
Most agents drown in context or starve without it. Dhee does the boring hard part in the middle.
- Reads summarize files instead of dumping them.
- Searches return compact matches with expansion pointers.
- Tool output becomes reusable evidence, not one-turn sludge.
- Task contracts define files, constraints, tests, and proof obligations.
- SceneCards become the default retrieval object.
- Raw transcripts stay out of prompts unless explicitly requested.
The model gets the part of the world it needs, in the form it can use.
This is the new Dhee core.
Series -> Season -> Episode -> Scene -> SceneCard -> MemoryItem
In plain English:
- Series is the biggest purpose. Example: become a successful CTO.
- Season is a period of the story: learning, struggle, first success, downfall, comeback.
- Episode is the meaningful arc of a day.
- Scene is a bounded work moment with a hero, intent, action, obstacle, result, and outcome.
- SceneCard is the retrieval card the agent sees later: compact, evidence-backed, privacy-aware.
- MemoryItem is promoted only when the SceneCard is worth keeping long-term.
That story gives the model anticipation. The prior is advisory, never bossy: explicit user intent, facts, privacy, and proof gates win.
Dhee can run model-free, but the production retrieval path is built for serious agent work:
pip install "dhee[nvidia,zvec,mcp]"
dhee key set nvidiaCurrent high-quality stack:
- Embedder:
nvidia/llama-nemotron-embed-vl-1b-v2 - Reranker:
nvidia/llama-3.2-nv-rerankqa-1b-v2 - Vector backend:
zvecthroughdhee-accel - Routine narrative rollups:
google/gemma-4-31b-it - Series-level strategic rollups:
moonshotai/kimi-k2.6
Deterministic filters still go first: secrets, private scenes, contradicted cards, superseded cards, and code-mutation proof gates cannot be talked around by a high similarity score.
Add Dhee to any MCP client:
{
"mcpServers": {
"dhee": { "command": "dhee-mcp" }
}
}Scene intelligence tools:
dhee_scene_start
dhee_scene_event
dhee_scene_end
dhee_scene_context
dhee_narrative_prior
Core context tools:
dhee_context_bootstrap
dhee_read
dhee_grep
dhee_bash
dhee_context_pack
dhee_scene_search
A good agent loop is simple:
bootstrap -> start scene -> gather evidence -> retrieve context -> act -> verify -> end scene -> save digest
Ask: "If I touch this file, what breaks?"
Dhee's repo brain answers with graph-backed evidence:
- impacted files and symbols
- impacted routes and React components
- likely tests to run
- owners from git history
- related failure signatures
- source windows with line numbers
The repo brain is git-SHA scoped and persisted under .dhee/context/repo_brain/, so the agent does not rebuild the same understanding every session.
Normal memory says: "Here are some old notes."
Dhee says:
- this is the user's bigger goal
- this is the current season of work
- this is today's episode
- this scene has these constraints
- these cards are safe to use
- these cards were rejected and why
- this action needs proof before mutation
The result is an agent that feels less random because it can see the arc.
On LongMemEval full 500-question recall:
| System | R@1 | R@3 | R@5 | R@10 |
|---|---|---|---|---|
| Dhee | 94.8% | 99.0% | 99.4% | 99.8% |
Reproduction notes and outputs live in benchmarks/longmemeval/.
The open source package includes the local memory OS, context compiler, DheeFS, MCP servers, CLI, UI, runtime daemon, handoff bus, repo brain, update capsules, narrative scene intelligence, release checks, and proof tooling.
Team governance, hosted dashboards, org policy, and managed source connectors can build on top of these local primitives. Dhee itself is useful without a hosted account.
pip install -e ".[dev,nvidia,zvec,mcp]"
pytestFocused release checks:
python -m pytest tests/test_narrative_scene_intelligence.py tests/test_mcp_tools_slim.py tests/test_scene.py tests/test_temporal_scenes.py tests/test_reranker_defaults.py tests/test_nvidia_embedder.py -q
python -m build
python -m twine check dist/*MIT. Built by Sankhya AI Labs.