Hack Hydra 2026 · Track 3

Agent memory that knows
what changed — and when.

Every fact is a claim with provenance and a validity window. Overwrites and contradictions are first-class graph structure. A graph-probe router answers cheaply, escalates conflicts, and abstains when the graph can't back an answer.

100%accuracy (probe arm)
25benchmark questions
8conflict scenarios
~71msp95 latency

Vector memory doesn't track truth.

When deadlines move, owners hand off, and status changes across Slack, Linear, and meetings — flat chunk stores retrieve stale text. They can't tell you what's current, what was believed last week, or when to say "I don't know."

Flat vector RAG
  • Retrieves stale chunks alongside current ones
  • Can't detect cross-source contradictions
  • Guesses on unanswerable questions
  • Re-derives state at read time — slow, fragile
24% accuracy
HydraClaim
  • SUPERSEDES edges track every overwrite
  • CONTRADICTS edges flag unresolved conflicts
  • Typed-coverage abstention — exact, not threshold
  • Bitemporal queries — "what was true at T" is a filter
100% accuracy

Memory as a graph, not a haystack.

Supersession chains

New facts create SUPERSEDES edges — history is never overwritten. Reconstruct any fact's full timeline with bounded path traversal.

Conflict detection

When Slack says one thing and Linear says another, CONTRADICTS edges mark the disagreement. Trust scoring arbitrates.

🎯

Typed abstention

If no claim exists for the asked (subject, predicate), the system declines and reports what it searched. No guessing.

Bitemporal queries

"What was believed as of last Tuesday?" is a property filter on valid_from/valid_to — a Cypher query, not inference.

🔀

Graph-probe routing

2–3 bounded Cypher queries measure coverage, conflicts, and chain depth. Clean → FAST. Messy → DEEP. Empty → ABSTAIN.

📎

Cited answers

Every answer traces back to verbatim quotes, authors, and sources. Provenance is structure, not prompt engineering.

From messy events to cited answers.

1

Ingest

Slack, Linear, meeting transcripts → LLM extractor pulls atomic claims with a closed predicate vocabulary (~12 types).

2

Reconcile

Deterministic rules detect overwrites (SUPERSEDES) and conflicts (CONTRADICTS). Deduplication. No ambiguity.

3

Store

Claims, evidence, entities, and typed edges land in HydraDB via batched UNWIND writes. Idempotent.

4

Query

Classify → graph probe → route FAST / DEEP / ABSTAIN → cited answer with timeline and provenance.

Ablation on 25 questions, 8 scenarios.

Arm Accuracy Knowledge update Abstention P / R Queries/q p95 latency
Naïve RAG24.0%100%0% / 0%0.96 ms
Question Router76.0%100%100% / 50%4.8111 ms
Always Deep84.0%100%100% / 50%5.084 ms
Router + Graph Probe100%100%100% / 100%4.771 ms

The naïve RAG baseline picks the single active claim with the most word overlap. It cannot see supersession chains, cannot surface conflicts, and guesses on every abstention question. The graph probe gives HydraClaim precise, typed coverage.

Ask the memory graph.

Connected to a live HydraDB instance with 8 pre-ingested scenarios. Ask about deadlines, owners, conflicts, or things that were never recorded.

connecting…
Memory graph
entity active claim superseded SUPERSEDES CONTRADICTS