Vectadyne Research

The most advanced memory system ever built for engineering agents.

Most "agent memory" is a vector store with a similarity search bolted on. Vectadyne is a governed control plane: a typed, bi-temporal, provenance-complete knowledge graph with a deterministic trust boundary between having a memory and acting on it. This page is how it works.

3-axis memory model Bi-temporal validity Deterministic admission Hash-chained provenance
Abstract

A fleet of agents generates enormous experience and almost never reuses it safely. We treat memory as five logical planes — evidence, canonical graph, guides, retrieval + admission, governance — where evidence is immutable and everything downstream is derived and traceable back to it. The result is experience reuse that is shared, explainable, and governed: one agent's failure becomes the whole org's guardrail, without a rulebook anyone has to maintain by hand.

01 · The memory model

Three orthogonal axes, not a flat embedding.

Every memory node is a point in a three-dimensional space. Because the axes are independent, the system can answer who a memory is for, when it is true, and what shape of knowledge it is — separately and precisely.

SCOPE — who for?

A governance-inheriting lattice; a policy at org is enforced everywhere below it.

org › domain › team › repo ›
env › role › agent › tool
TIME — when true?

Bi-temporal: when a fact was valid in the world vs. when the system asserted it.

valid_from / valid_to
asserted_at / review_after
KIND — what shape?

Seven typed kinds with six edge types — a real graph, not a bag of chunks.

fact decision lesson policy skill guide evidence
02 · Time & forgetting

Memory that forgets on event, not on a timer.

Naïve decay throws away knowledge that is still correct. Vectadyne forgets for a reason — each transition is a recorded, auditable event, and the bi-temporal record means you can always time-travel to what an agent knew at decision time.

active

Current and trusted — admitted at full weight.

disputed

Contradicted by newer evidence; surfaced, but down-weighted.

stale

Past its review_after horizon; flagged for re-derivation.

deprecated

Superseded or retracted; kept for audit, excluded from retrieval.

03 · Retrieval → admission

A deterministic trust boundary between having a memory and acting on it.

Retrieval is cost-adaptive: cheap lexical and scope filters first, dense and graph expansion only when the budget and the query warrant it. Then admission — the part that matters — gates candidates deterministically, on both the write path and the read path.

01
Scope + lexical filter
Narrow to the querying agent's scope lattice and cheap keyword matches — no vector cost yet.
02
Cost-adaptive expansion
Dense retrieval and 1-hop graph expansion fire only when the budget and query complexity justify them.
03
Write-time admission
Before a derived memory is promoted, provenance and confidence thresholds must be met — the anti-poisoning boundary.
04
Read-time admission
At preflight, the resolution walk resolves policies org→narrow and returns allow / warn + cited provenance.
04 · Provenance

Every memory traces to hash-chained evidence.

The evidence log is append-only and hash-chained — a tamper-evident system of truth. Canonical memory is derived, never hand-typed, so any node, guide, or verdict can be walked back to the raw traces, PRs, and CI runs that produced it.

evidence ev-8842 ⛓ sha:9f3c…a10
↳ derives lesson led_1421
↳ gates preflight #4471
05 · Authority

Trust is computed, not assumed.

A node's weight at admission combines several priors, so a fresh, corroborated, high-scope lesson outranks a stale, disputed one — automatically.

confidence
freshness_prior
scope weight
corroboration
06 · Taxonomy

We subsume the standard memory taxonomy.

The vocabulary buyers and papers use — working / episodic / semantic / procedural — maps cleanly onto our seven kinds. We don't replace the taxonomy; we make it typed, scoped, and governable.

Standard taxonomy
Vectadyne kinds
Working memory
Lives in the agent runtime — we feed it resolved guides + admitted citations
Episodic memory
evidence + lesson (what happened, what failed)
Semantic memory
fact + decision (what is true, what was chosen)
Procedural memory
skill + guide + policy (how to act, what to require)
07 · Evaluation

A system that evaluates itself.

An internal eval + safety harness ships with the product as vec e2e, measuring failure-prevention against a held-out set of known regressions. Directional results from pilot repositories:

−79%
Repeated-failed-fix rate
82%
Preflight heed rate
−64%
Regression recurrence
2.4h
Time to first correct patch
Illustrative figures from private-beta pilot repositories; not a published benchmark.

Open research questions.

We publish our uncertainties, not just our claims. Each of these has a written trigger for when we'll adopt a heavier design.

v1 forgetting is event-driven. We'll add a continuous decay sub-score only if the eval harness shows stale-but-not-yet-flagged memory materially degrading verdict quality — the freshness_prior already carries the plumbing a decay term would slot into.

v1 is deliberately 1-hop. Multi-hop traversal, community detection, and GraphRAG-style community summaries are deferred until retrieval quality on real fleets demonstrably needs them — not before, because they cost latency on the preflight hot path.

Auto-derived lessons are both the only value source and the top attack surface. We require provenance-backed derivation and write-time admission thresholds, so a single crafted trace cannot mint an enforced memory.

See the hive mind on your own repo.

The research is only interesting if it prevents a real failure. Watch it warn on a repeated fix in your codebase.

Book a demo Back to vectadyne.io
Vectadyne © 2026 Vectadyne, Inc.
Home Model Evaluation