Govern what enterprise AI agents are allowed to remember.
Authoritative memory persists. Stale, rejected, and rolled-back records remain auditable without silently shaping future answers.
For regulated deployments, FieldHash can run against customer-approved models and infrastructure, with exportable evidence logs and PQC-capable provenance where configured.
Governed learning
A governed control plane for memory influence.
FieldHash promotes authoritative memory into governed state, holds the current fact under stale-context pressure, survives rollback and compaction, and logs audit telemetry when a control fails. Enterprise pilots start with stale-context suppression, rollback trails, approved-memory lineage, and exportable evidence logs. The studies below are bounded supporting evidence under documented test conditions, not external validation.
Primary synthesis
Governed continual-learning loop
The full arc: promotion identifies authoritative memory, arbitration keeps it intact, lifecycle state preserves rollback and compaction, and audit controls catch stale or corrupted influence.
Compliance review
Show which authoritative memory entered the answer path, which stale or rejected records were blocked, and which audit artifacts bind the event.
Long-lived agents
Keep authority stable across updates, rollback, repair, compaction, and repeated reads.
Research workflows
Preserve decisions, caveats, null results, and follow-up plans without turning every note into a hidden prior.
Supporting evidence studies
Authoritative memory
Automatic memory promotion
The current record is selected, promoted, and governed before generation.
Open supporting studyStateful learning
Governed learning lifecycle
Authoritative state survives update, rollback, repair, compaction, and repeated reads.
Open supporting studyMemory arbitration
Governed memory pressure
Reviewed decisions survive adversarial stale context.
Open supporting studyFalsifiable controls
Auditability diagnostic
Deliberate governance failures are detected instead of silently steering answers.
Open supporting studyHow it integrates
A governance layer between memory and the model.
FieldHash does not require teams to replace their agent, model, or vector store first. It sits on the answer path: inspect candidate context, enforce approved state, suppress stale or rejected influence, then emit an audit packet your reviewers can inspect.
ACTIVE GATE PIPELINE SIMULATION
1. Dispatch Query
The client application initiates a standard operational check, sending a query to the agent gateway under ordinary VPC or on-prem connectivity.
[Client] Query dispatched: "What is Project Alpha codeword?"
Most AI memory cannot show what governed the answer.
Standard AI can produce a strong answer. Once memory enters the workflow, teams need to know which facts, corrections, and decisions were allowed to steer the next one.
FieldHash preserves what survives review, while stale, rejected, or rolled-back state remains visible for audit without silently becoming a future prior.
Standard workflow
Retrieve, append, repeat.
The next answer depends on whatever context retrieval surfaces unless a human manually carries status, caveats, and prior decisions forward.
FieldHash workflow
Review, govern, verify.
Validated context and contradiction resolutions become inspectable future influence; rejected and rolled-back state stays stored without steering the answer.
Why not just use RAG?
You probably already are — and FieldHash runs on the same retrieval substrate. It is not a better retriever and not a more accurate model: given the same context, answer quality is comparable. What it adds is the part RAG leaves to chance — control over which memory is allowed to influence an answer, and a record of what did.
Influence control
Retrieval surfaces whatever looks similar. FieldHash gates it: approved, current state shapes the answer, while stale, rejected, superseded, or rolled-back state is blocked from live influence.
Auditability
A tamper-evident record of what was approved, rejected, or rolled back — and what actually entered the answer path — so a reviewer can confirm it rather than take it on trust.
Reversibility
Retract a decision and it stops steering future answers. Plain retrieval keeps surfacing a superseded fact because it still looks relevant; governed state does not.
Memory stores facts. Governance controls influence.
The underlying LLM stays frozen during normal use. Learning lives in scoped memory, routing hints, symbolic state, contradiction handling, hub compression, and promotion rules that decide what can shape later answers. Enterprise deployments can configure the layer against customer-approved private, local, VPC, or on-prem models when the work requires tighter data boundaries, with governed-memory events bound to audit artifacts where configured.
Input
A question, document set, dataset, or research session starts the loop.
Retrieve
Candidate memories and prior artifacts surface by reasoning context.
Gate
Scope, evidence, relevance, novelty, and governance checks filter influence.
Answer
The model responds with approved context and current task constraints.
Log
State shifts, caveats, sources, and outcomes become inspectable artifacts.
Promote
Only useful, stable signals become durable context for future sessions.
What governance prevents: rejected brainstorms, stale corrections, and unrelated project notes should remain auditable without becoming hidden priors.
Read the architectureGoverned Memory enables Deep Synthesis.
Standard deep research reports what the sources say. Deep Synthesis uses governed memory to keep hypothesis lineages, caveats, null results, and validation plans inspectable across a research thread. It is an application of the same control plane, not the primary enterprise purchase path.
Hypothesis-first synthesis
New explanations stay linked to evidence, caveats, confounders, and source context.
Null-result aware
Weakened paths are not re-promoted unless the changed discriminator is explicit.
Validation-routable
Strong ideas are paired with controls, readouts, and explicit failure conditions.
Research Lab inside
The validation lane runs tournaments, preserves negative results, and returns interpretable structure when supported.
Synthesis case studies
Applications of the synthesis loop.
Emergence synthesis
Noisy signals reduce into reusable research priors.
ReadRegulatory synthesis
Compliance analysis with caveats and audit posture.
ReadBusiness synthesis
Hidden risk surfaced with readable rationale.
ReadBiology synthesis
A tauopathy mechanism hypothesis shaped into testable research lanes.
ReadValidation lane
Research Lab is the test bench inside synthesis.
When source material includes executable data, Deep Synthesis can escalate from explanation to experiment: competing model families, falsification plans, held-out checks, and retained negative results.
Four governed-memory evidence points.
These are the studies behind the governed-memory wedge: automatic promotion, stateful lifecycle persistence, memory arbitration, and falsifiable audit controls. Tool-context compression remains a supporting infrastructure diagnostic, not the core claim.
Start with the loop case studyAutomatic promotion
The system governs the authoritative memory.
In internal automatic-promotion diagnostics, FieldHash identified the authoritative memory and recovered 90/90 exact current tokens across Gemini 3.5 Flash, Claude Opus 4.7, and GPT-5.5 on a Claude-authored disjoint n=30 corpus, while retrieval-only and prompt-only smart-memory controls recovered 36/90 and 40/90. A same-budget Gemini two-pass smart diagnostic on the same corpus selected the current record 30/30 and answered 28/30 with zero stale substitutions, narrowing the claim to governed, auditable answer-path control rather than basic semantic selection. On same-family n=100 provider replications, Gemini 3.5 Flash and GPT-5.5 each reached 100/100 with zero stale-token mentions; Claude Opus 4.7 reached 95/100, with the remaining misses caused by empty provider responses rather than stale substitutions. In a provider-sensitivity fact-extraction audit on the same n=100 corpus, Gemini reached 99/100 role-equivalent current facts and 68/100 exact spans, while GPT-5.5 reached 95/100 and 76/100; strict source-span fidelity and provider-invariant extraction are not claimed as solved.
Stateful lifecycle
The governed state survives rollback.
In an internal n=200 lifecycle diagnostic, FieldHash preserved governed state across update, rollback, repair, compaction, and repeated reads while stateless LLM selectors missed rollback.
Governed memory pressure
The reviewed answer survives stale context.
In the refreshed May 23 internal seeded-authority memory-pressure benchmark, the same frontier LLMs with FieldHash governed memory enforced reviewed/current state and recovered the approved-current fact in 600/600 cases across Gemini, Claude Opus, and GPT provider paths. Retrieval-only memory without FieldHash governance metadata recovered 415/600. The governed path reduced mean memory context exposed to the LLM to 2.00 of 10 retrieved candidates before answer construction, versus all 10 candidates in the retrieval-only baseline. Across the same three provider paths, adding a prompt-only instruction to prefer current/reviewed records improved the baseline to 464/600, but still left 136 stale-context failures and exposed all 10 retrieved memories. This supports governed-state enforcement under stale-context pressure, not a claim of superior authority inference from raw text.
Audit controls
The audit layer catches broken governance.
The governed memory auditability diagnostic passed 437/437 deterministic checks across 36 lifecycle scenarios. That includes 257 positive governance invariants and 180/180 negative controls that deliberately disabled governance or corrupted state, confirming the suite catches stale exposure, rejected-context promotion, missing superseded_by links, scope leakage, and stale re-promotion.
Continuity you can inspect.
Companion now lives on its own product surface — a working demonstration of governed memory: prior decisions return, corrections supersede stale facts, and continuity stays inspectable.
Visit companion.fieldhash.ai
Govern what influences the answer — and verify the record.
FieldHash
The verifiable audit layer of governed memory. Governance events are hash-chained, signed into checkpoints and FieldHash certificates, and transparency-anchored so an auditor can confirm the record of what was approved, rejected, or rolled back was not quietly changed. Enterprise deployments can support customer-owned audit logs, EU-region cloud, customer VPC, on-prem infrastructure, and SIEM/GRC export where configured. Post-quantum where configured; operator-resistant when the anchor is held outside the operator boundary; tamper-evident, not tamper-proof.
Evidence: FieldHash adaptive spoofing campaign.
FieldHash overviewThe governance engine
The Gnosis engine decides what may influence an answer and keeps bounded adaptation governable — static review, isolated testing, coherence checks, signed decisions, and rollback before any sensitive change reaches production.
Gnosis overviewNeed AI memory
your reviewers can trust?
Start with the governed-memory evidence, then request technical review for architecture notes, benchmark methodology, ablation summaries, and trace examples.
The whitepaper remains available for technical readers; enterprise evaluation starts with memory governance.