Hack Hydra 2026 · Track 3, memory and context retrieval
Lacuna
Temporal, provenance-first memory for AI agents. It remembers what changed, preserves what was true, exposes disagreements and refuses to invent an answer when the evidence is missing.
The problem
Long-running agents fail when memory goes stale, contradicts itself or loses the evidence it came from. A model can reason correctly from a premise that stopped being true months ago, and four agents that each keep their own memory slowly become four versions of reality.
How it works
Lacuna is an evidence-bearing memory layer on HydraDB. Every claim stays tied to the sentence it came from; a correction is linked to what it replaced instead of overwriting it; sources that disagree are shown as a disagreement; and when the evidence is missing, the answer says so.
Ask in plain English and it returns the current answer, its source, its timeline and any conflicts. The same shared memory is exposed through the web workspace, a CLI, an API and seven public read-only MCP tools, so a compatible AI client gets a focused evidence pack instead of the whole history. The final release carried 453 graph nodes and 682 relationships with no orphan edge, a 64-question evaluation that answered identically over MCP stdio, MCP over HTTP and the CLI, and a verified product film.
Highlights
- Temporal claimsNew information links to what it superseded; the history stays queryable.
- Exact provenanceEvery answer opens to the sentence it was drawn from.
- Contradictions surfacedSources that disagree are shown as a disagreement, never averaged away.
- AbstentionNo evidence, no answer: the system says what was never stated.