Exact
Identifiers and clauses
Part numbers, standard names, customer codes, and exact quoted language need lexical retrieval.
ZERO1 Core research / Evidence memory
Memory should know what it is allowed to prove.
Evidence Memory is a supporting capability inside ZERO1 Core, shaped by a broad arXiv research review. The search identified 200 candidate papers across primary and supplemental searches. The recorded primary query returned 101 candidates, and 42 directly relevant papers were synthesized into architecture decisions around source, authority, time, permissions, correction, and retrieval.
01 / Source
Permissioned records enter intact.
Documents, email, prior work, policies, and live systems retain source identity before any memory write is proposed.
Research synthesis / Source-linked architecture
200 identifiedcandidate papers across the broader arXiv search
101 recordedcandidates returned by the reproducible primary query
42 synthesizeddirectly relevant papers in the working registry
Human-gatedauthority remains with the person who owns the decision
Designed from a reproducible, query-bounded review. The research points to a source-linked architecture that separates retrieval, authority, current state, and release.
Research basis / Official arXiv records
01 / The finding
Different questions need different memory.
The best memory is not one database.
A company does not remember in one format. Literal identifiers, equivalent prior work, connected dependencies, and current operating state are different retrieval problems. Treating them as one problem creates confident gaps.
Semantic
Equivalent prior work
Similar projects may use different wording. Dense retrieval finds useful analogs without a literal match.
Relational
Connected evidence
Products, jobs, rate lines, policies, and decisions need typed relationships and multi-step traversal.
Current
Live operating state
Inventory, status, approvals, and current rates belong in approved system queries, not stale summaries.
Retrieval is only useful when the result can show where it came from, when it applied, and whether it was allowed to govern the work.
02 / Approach comparison
Method classes, not vendor scorekeeping.
Every approach has a useful operating range.
The useful question is not “graph or vector?” It is which method should answer this request, under this authority and time boundary, with evidence strong enough for the next action.
| Approach | Works well when | Breaks down when | Role inside Core |
|---|---|---|---|
| Full allowed context | The source set is small and the relevant evidence fits cleanly in context. | History grows, irrelevant material distracts the model, or cost and latency rise. | Controlled baseline and bounded fallback, not the default memory layer. |
| Exact / BM25 | The request contains literal identifiers, clauses, codes, names, or quotations. | The same idea is expressed with different wording or spread across relations. | First route for exact requirements and source-span recovery. |
| Dense vectors | Semantic similarity and paraphrased prior work matter more than exact wording. | Authority, time, permissions, identity, and multi-step dependencies determine the answer. | Candidate generation for analogs, followed by evidence and policy filters. |
| Graph retrieval | Entities and evidence form useful paths across projects, decisions, rules, and records. | The graph is incomplete, an extraction is wrong, or centrality is mistaken for authority. | A typed, source-linked projection over versioned assertions, not the sole record. |
| Fixed hybrid | A stable blend improves coverage across a known workload. | A fixed blend can add avoidable cost or mask task-specific failures. | A reference baseline for controlled comparison. |
| Adaptive evidence routing Core architecture | The system can classify the evidence need and route only to allowed, relevant stores. | The router, write policy, or evidence filters are not independently tested. | Exact, sparse, dense, graph, file, and live routes under one source-and-authority contract. |
This is an architectural comparison, not a performance leaderboard. The central finding is that retrieval methods should be assigned by evidence need, source authority, time boundary, and allowed use.
03 / Architecture consequence
Research translated into operating rules.
An evidence system with memory inside.
The architecture preserves the source, stages each memory write, retains history, attaches authority and access, retrieves original evidence, and stops consequential work at a named human gate.
01
Approved sourcesDocuments, mail, prior work, policies, and read-scoped systems.
02
Immutable evidenceVersion, checksum, evidence span, tenant, access scope, and retention.
03
Staged memory writeEntity resolution, conflict scan, authority class, and policy checks.
04
Versioned assertionsValid time, recorded time, lifecycle state, and derivation history.
05
Evidence routerTask-specific retrieval after permission, authority, version, and time filters.
06
Result + receiptSupported fields, conflicts, missing inputs, checks, and reviewer disposition.
Adaptive evidence routes
The work selects the route. A fashionable database does not.
Exact / BM25
Dense vector
Typed graph
File search
SQL / API
A generated summary never becomes authoritative merely because it was stored. Critical assertions remain bound to their evidence and allowed use.
Human controls price + release
04 / What the papers changed
Primary sources linked. Conclusions remain bounded.
Six lessons worth building around.
The review did not produce a universal winner. It exposed where simple memory claims fail once the system has to act, route across source types, preserve authority, survive corrections, and remain comparable.
01 / Action
Recall is not operational reliability.
MemoryArena reports that systems near saturation on a conversational benchmark can still perform poorly when memory must guide later actions. Core therefore evaluates memory through the grounded work product and the action it supports, not chat recall alone.
Primary source / arXiv:2602.16313 ↗02 / Routing
Choosing the right source is not enough.
WorkSurface-Bench reports 98.7 to 99.8 Route F1 under gold-constrained tools while answer accuracy remains 56.1 to 75.3%. Core treats source routing and final evidence use as separate problems.
Primary source / arXiv:2607.25765 ↗03 / Domain
Professional work needs domain-aware structure.
IFCMemoryBench combines prior sessions with live engineering-model queries. Its strongest tested system reaches 32.4% answer accuracy under deployment-realistic ingestion, exposing a gap between topical retrieval and usable project knowledge.
Primary source / arXiv:2607.26072 ↗04 / Authority
A stored claim is not automatically authorized truth.
AuthMem-Bench observes authority collapse in 48 of 49 tested configurations. Core’s design keeps the source class and permitted use attached through consolidation and retrieval.
Primary source / arXiv:2608.01679 ↗05 / Lifecycle
Memory is a sequence of state changes.
MemOps models remembering, forgetting, updating, and reflecting as explicit operations with a trigger, target, scope, state transition, and evidence. Core’s design retains supersession and correction instead of silently overwriting history.
Primary source / arXiv:2607.12893 ↗06 / Evaluation
Change one variable at a time.
MemDelta shows that swapping only the embedding model can shift accuracy and reverse architecture rankings. Core treats the model, embedding, retrieval budget, sources, and output schema as controlled variables when comparing methods.
Primary source / arXiv:2606.29914 ↗05 / Review method
Reproducible enough to challenge and extend.
AI accelerated the review. The method stayed inspectable.
AI accelerated search and synthesis across official arXiv records. The query, scope rules, screening method, linked sources, and architecture decisions remain visible so the reasoning can be challenged.
01 / Query
Search the official index.The primary arXiv API query established a reproducible starting set before supplemental exact-name searches.
02 / Screen
Apply written scope rules.All primary candidates received title screening; directly relevant papers received abstract and metadata screening.
03 / Synthesize
Map findings to the job.The working registry spans recall, action, routing, time, authority, safety, cost, and professional transfer.
04 / Bound
Keep claims inside the sources.Each public conclusion points back to the paper or operating principle that changed the architecture.
Primary arXiv query((all:"agent memory" OR all:"long-term memory") AND (ti:benchmark OR ti:bench OR ti:evaluating OR ti:evaluation))
06 / Governed improvement
Reviewed outcomes become controlled changes.
Improve from reviewed outcomes, not from its own guesses.
Core compares a prepared work product with the reviewer’s disposition, records the correction and its evidence, stages a bounded change, replays it against approved cases, and keeps promotion behind an authorized person.
01 / Prepare
Build the result.
Prepare a source-linked work product under a versioned Result Spec.
02 / Review
Record the disposition.
An authorized reviewer accepts, rejects, or corrects the work and records why.
03 / Propose
Stage a bounded change.
Create a candidate update to the Result Spec, source mapping, retrieval route, or check.
04 / Replay
Test before promotion.
Run accepted and sealed cases, including authority, permission, and exact-retrieval regressions.
05 / Promote
Keep a human gate.
An authorized person approves a new version; the previous version remains available for rollback.
Evidence
Corrections stay linked.
Every accepted correction retains the source and reviewer disposition that justified it.
Replay
Changes face prior cases.
Candidate updates run against approved examples before they can affect future work.
Authority
Promotion stays human-controlled.
An authorized person approves the new version, and the prior version remains available for rollback.
07 / Selected reading
Official arXiv records linked directly.
Start with the papers that changed the design.
These primary sources produced the clearest architecture decisions around retrieval, action, authority, correction, and controlled comparison.
2608.15008Holistic memory-substrate evaluationNo single substrate dominates across tasks, scale, and operating regimes.↗
2607.26072IFCMemoryBenchProfessional domain transfer using prior sessions and live structured queries.↗
2607.25765WorkSurface-BenchSource routing and answer quality across documents, tables, and graphs.↗
2608.01679AuthMem-BenchSource authority and allowed use must survive memory consolidation.↗
2607.12893MemOpsRemember, forget, update, and reflect as explicit state transitions.↗
2606.29914MemDeltaControlled baselines and hidden confounds in memory evaluation.↗
2602.16313MemoryArenaMemory tested through interdependent actions, not recall in isolation.↗
2410.10813LongMemEvalExtraction, multi-session reasoning, time, updates, and abstention.↗
ZERO1 Core research
See how the research appears in a working interface.
The RFQ specimen shows a source-linked quote draft with visible gaps and estimator approval.