Structural + semantic index
The graph holds explicit relationships alongside embeddings, so precise and fuzzy questions both land.
- ↳project graph
- ↳hybrid retrieval
- ↳entity + relation extraction
havn systems · intelligence layer
havn-lore.
Persistent structural and semantic memory for projects, agents, and the work around them.
A memory service that keeps context attached to the material, so people and agents can retrieve it across sessions, repos, and surfaces.
package
havn-lore · Rust + MCPhavn / havn-lore connected
havn-lore is the layer that remembers structure. Where Knowing makes a single canvas queryable, havn-lore holds the durable graph across projects: what a thing is, how it relates to the rest, and why it mattered when it was written down.
It is built as a service with a typed boundary. Agents and product surfaces ask it questions through MCP; it answers from indexed material instead of from a chat history that scrolls away.
See it in the architectureAgents forget, and so do teams.
Most AI tooling starts every task from a cold context window. The result is confident work built on a partial picture: the same file re-read, the same decision re-litigated, the same constraint missed.
havn-lore keeps the picture warm. Structural and semantic memory are indexed once and served back with provenance, so the system can act on what is actually true about a project rather than what fits in a prompt.
Small boundary.
Large surface area.
Context that persists
Decisions, entities, and relationships outlive the session that created them.
Grounded answers
Retrieval carries provenance, so an agent can show its work instead of asserting it.
One memory, many callers
People and agents query the same substrate through a stable interface.
MCP as the boundary
Memory is exposed as typed tools, not a bespoke API per consumer.
- ↳MCP tool surface
- ↳typed queries
- ↳agent-ready contracts
Verification built in
An honesty guard checks answers against the indexed source before they leave the service.
- ↳honesty-guard invariant
- ↳offline eval harness
- ↳provenance on every result
How the layer
moves.
Every surface in Havn has a job in the loop. This one keeps its responsibility clear, then hands the work to the layer that comes next.
Ingest
Take repos, notes, and project material into the index.
Structure
Extract entities, relationships, and embeddings into one graph.
Serve
Answer typed questions over MCP with provenance attached.
Verify
Check the answer against the source before it is returned.
Nothing here
stands alone.
havn-lore is useful because it has a precise relationship with the other layers. Follow the handoffs and the product gets easier to understand.
memory substrate↗
Knowing
Handles per-canvas retrieval; havn-lore extends it to durable, cross-project memory.
knowing canvas↗
Board
Asks havn-lore for context while a person works, then keeps the answer in the surface.
compute foundation↗
Kernel
Provides the shared identifiers that let memory point back at real objects.
Give the system a memory it can be held to.
havn-lore is the intelligence layer beneath Havn: durable structure, grounded retrieval, and a boundary that agents and people share.