Spector

1. Executive Summary

Spector is a Java 25 memory engine for agents. It stores each memory as an off-heap "engram" — a quantized vector with importance, valence, decay and two flag bytes — and ranks recall with a fused SIMD score over similarity, importance and a power-law decay, widened by BM25, SPLADE and a Hebbian and entity graph. Beside the engrams sits a temporal knowledge graph of subject-predicate-object facts with valid time and transaction time. It runs embedded, as a standalone MCP server, or as the Synapse server with REST, gRPC and per-account namespaces.

What is notable is the deletion vocabulary and one working correction mechanism. memory_forget tells the agent that the bytes remain on disk and names memory_purge, which zeroes the payload and lists the copies it cannot reach. A background consolidator asks an LLM whether two near-duplicate memories contradict each other, flags the older one, and default recall skips it; a committed test asserts exactly that.

What is weak is the fact plane and the declared-but-unwired governance. Valid time is written in epoch seconds and compared in milliseconds. The fact read path never excludes a retracted fact. A "fail-closed" release gate filters on retracted, unverified and restricted flags that nothing in the tree sets.

The licence is Apache-2.0. The NOTICE adds that anyone using Spector "MUST provide visible attribution", with "Powered by Spector" text and a repository link, and withholds permission to offer it as a SaaS product under another brand. A reader deploying Spector should read both files.

Three marks: trust_state on the contradicted flag, bitemporal on the fact log with the unit mismatch stated as its limit, and negative_eval on two committed exclusion cases with positive controls. Section 9 names the four withheld.

2. Mental Model

Spector holds two kinds of belief and treats them differently.

An engram becomes a belief when it is written. remember embeds the text, writes the record and indexes the id, with no candidate state. The id is the identity: DedupGuardRelay refuses a second write under an id already in the index (DedupGuardRelay.java:41), and nothing compares content at the door. The source (USER_STATED, OBSERVED, INFERRED and six more) is a caller argument carrying a declared weight (MemorySource.java:34-77), not a gate.

An engram stops being a belief in five ways. forget sets the tombstone bit and removes the id from the index, so the same id can be written again (DefaultSpectorMemory.java:1132-1134). purge also zeroes the payload. Consolidation can mark it contradicted, or merge it with a duplicate into a new cns- memory and tombstone both sources (AbstractConsolidator.java:233-236). suppress hides it from recall until the process ends. Decay lowers its rank but, by the README's formula, never below a floor.

The contradicted flag is the only status the system moves. After a SEMANTIC or PROCEDURAL write, EagerConsolidator scans the tier for records within an L2 threshold and asks an LLM whether each pair contradicts (EagerConsolidator.java:229-296). On a YES, CadpContradictionResolver decides by timestamp, then storage strength, then id — newer wins — and sets FLAG_CONTRADICTED on the loser (CadpContradictionResolver.java:68-117). Default recall skips it.

Simulated and dreamed traces are a genre, not a status. Dream and constructive-simulation relays write records with FLAG_SIMULATED or FLAG_DREAMED, and recall withholds them unless allowSimulated is set. The flag is fixed at write.

A fact is a row in an append-only log. assertFactWithSupersession retracts every active fact with the same subject and predicate by appending a retraction row, then appends the new fact (TemporalKnowledgeGraph.java:348-378). A retraction never mutates the original. Whether a retracted fact stays hidden depends on the reader, and no production reader asks.

%% caption: how an engram and a fact become beliefs in Spector, and where each stops being one
flowchart TD
    W["memory_remember / REST / SDK"] --> D{"id already<br/>in index?"}
    D -- yes --> SKIP["write skipped"]
    D -- no --> E["engram written,<br/>consolidation flags = 0"]
    E --> R["default recall"]
    E -- "SEMANTIC or PROCEDURAL<br/>and an LLM provider" --> EC["EagerConsolidator:<br/>neighbours within L2 threshold"]
    EC --> J{"LLM reply<br/>contains YES?"}
    J -- no --> E
    J -- yes --> C["older record:<br/>FLAG_CONTRADICTED"]
    C -. "skipped unless<br/>includeContradictions" .-> R
    C --> RF["every fact about the loser's<br/>entities retracted"]
    E -- "memory_forget" --> T["tombstone bit,<br/>id freed, bytes remain"]
    T --> P["memory_purge:<br/>payload zeroed"]
    G["release gate: RETRACTED,<br/>UNVERIFIED, RESTRICTED"] -. "no writer<br/>in the tree" .-> R
    F["memory_fact_assert:<br/>valid_from in seconds"] --> L["fact row: validFrom, validTo,<br/>txTime in milliseconds"]
    L --> Q{"factsAbout: validAt<br/>compares milliseconds"}
    Q -- "validTo finite" --> X["excluded at every<br/>present-day instant"]
    Q -- "validTo open" --> Y["returned, retracted<br/>or not"]
    RT["memory_fact_retract"] --> RR["retraction row appended"]
    RR -. "no production read<br/>calls excludeRetracted" .-> Y

3. Architecture

The Maven reactor has four tiers. nucleus/ holds configuration, SIMD kernels, a GPU path and a vector index. memory/ holds spector-kernel, the off-heap storage layouts, and spector-memory, the cognitive engine. synapse/ holds the MCP module, the Spring Boot server, the CLI, batch import and export, and connectors. cortex/ is an Angular dashboard.

Persistence is Panama MemorySegment regions in memory-mapped files, one bundle per partition, under one directory per namespace, with a chunked write-ahead log that CheckpointEngine truncates after each checkpoint (CheckpointEngine.java:344). There is no external database; the NOTICE states that the core modules depend on the JDK alone.

The recall pipeline is a chain of relays assembled in RecallPathway (RecallPathway.java:333-359): query transduction, the governed release gate, the cortical tier scan, neuromodulatory scoring, graph expansion with temporal fact weaving, evidence fusion, lateral inhibition, BM25 and SPLADE fusion, RRF, sort-and-truncate with suppression, optional ColBERT rerank, MMR and temperature softmax. The AISME relays join the chain when spector.memory.aisme.enabled is true, which it is not by default (SpectorPropertyConstants.java:736).

Background work runs on Quartz: sleep consolidation, REM dreaming, default-mode wandering, homeostatic decay, graph enrichment every 30 seconds, index reconcile and checkpoint (QuartzMemoryScheduler.java:221-306).

Deployment and ergonomics

Embedded use needs JDK 25 with --enable-preview and the incubating Vector API module, and an embedding provider (Ollama, ONNX, OpenAI and others under memory/spector-providers). Contradiction detection and entity extraction need an LLM provider; without one they do nothing. The standalone MCP runner starts an in-process kernel or connects to a local daemon, and Docker Compose starts the server and dashboard. The store is binary and not repairable by hand; the batch exporter writes JSONL.

4. Essential Implementation Paths

Write. DefaultSpectorMemory.remember checks the MutationPolicy, chunks long text, embeds, and calls rememberPathway.ingestCognitive (DefaultSpectorMemory.java:453-490). The pathway runs DedupGuardRelay, tag transduction, a surprise relay, CorticalWriteTransactionRelay, synaptic graph linking and knowledge-graph enrichment. For SEMANTIC and PROCEDURAL writes it then submits the id to EagerConsolidator (:482-484).

Recall. DefaultSpectorMemory.recall → RecallPathway. The tier scan's SlabScanner skips tombstoned, contradicted and simulated records inside the scan loop (SlabScanner.java:75-98); SemanticRecallStrategy and RecallCandidateGatherer repeat the contradicted check (SemanticRecallStrategy.java:161-162, RecallCandidateGatherer.java:173).

Contradiction. AbstractConsolidator.evaluateAndResolvePair asks ContradictionDetector.areContradictory and, on true, calls CadpContradictionResolver.resolve (AbstractConsolidator.java:119-126). The detector returns false when no LLM provider is set and treats any reply containing YES as a contradiction (ContradictionDetector.java:45-47, :62). The resolver marks the loser, adds a TYPE_CONTRADICTS hyperedge, and retracts every fact about the loser's entities whose object is not one of the winner's (CadpContradictionResolver.java:163-180).

Forget and purge. forgetWithResult tombstones, appends a WAL FORGET, and removes the id (DefaultSpectorMemory.java:1114-1140). MemoryForgetTool reports a missing id as nothing forgotten rather than a success (MemoryForgetTool.java:48-61). purge zeroes payload and text bytes and reports graph references removed and copies not reached.

Facts. memory_fact_assert → DefaultSpectorMemory.assertFact → assertFactWithSupersession (DefaultSpectorMemory.java:2190-2204). REST GET /api/v1/memory/facts → factsAbout(entity, asOf) → factsAbout(entityId).validAt(asOf).resolve() (:2216-2227). During recall, TemporalFactWeavingStage appends [Temporal Fact: …] lines to matching results (TemporalFactWeavingStage.java:144).

MCP. SpectorToolRegistry.handlers registers 29 tools (SpectorToolRegistry.java:86-114); the README's table says 16. In Synapse, each call binds a namespace from the namespace argument through a deny check before the tool runs (McpServerConfig.java:249-255).

5. Memory Data Model

Scope is the directory. A namespace is a directory tree with its own bundles, created per account or per named namespace. There is no namespace key on a record and no scope predicate on a read. ADR 0062 states the security property directly: client-supplied namespace, workspace_id or agent_id "never change which memory is returned" (docs/adr/0062-memory-organization-three-plane-architecture.md:356).

The tool schemas say otherwise. memory_recall.json describes workspace_id as scoping recall "to the workspace's shared memory pool" with agent_id "for RBAC", and memory_remember.json makes the same promise for writes (memory_recall.json:71-78, memory_remember.json:69-75). No Java file outside tests reads either argument, so a model told it wrote to a shared pool wrote to its own namespace.

Provenance. ProvenanceMemory records which session turns a consolidated semantic memory came from (ProvenanceMemory.java:41-45). It is lineage for consolidation; forgets, suppressions and flag changes do not appear in it.

6. Retrieval Mechanics

Recall is automatic within the engine and tool-mediated for the agent. By the README, the core score is α·sim + β·I·D(t), with decay as a bucketed power law, after gates for the live bit, tag containment and a valence window. BM25 and SPLADE candidates are fused with the vector list by RRF, and graph expansion admits Hebbian neighbours at a discount, then temporal and entity hops.

Status filters sit in the scan. Contradicted records drop unless includeContradictions is set, which defaults to false (SpectorPropertyConstants.java:1215); the MCP recall tool does not expose the option. Simulated and dreamed records drop unless allowSimulated. Suppressed ids drop in SortAndTruncateRelay.

GovernedReleaseGateRelay drops a candidate whose consolidation flags carry RETRACTED, a RESTRICTED one when no persona is supplied, and an UNVERIFIED one below minTrustScore (GovernedReleaseGateRelay.java:58-81). The setters withRetracted, withUnverified and withRestricted exist (EncodingHeaderFields.java:454-468) and only tests call them. The production writes to that byte set CONTRADICTED, PURGED or, through the remember overlay, CRYSTALLIZED. The gate runs on every recall and has nothing to drop.

Temporal fact weaving reads tkg.factsAbout(entityId).validAt(asOf).resolve() (TemporalFactWeavingStage.java:144). Its display code guesses the unit — a value above 10^11 is read as milliseconds, anything else as seconds (:160-163) — while the filter above it does not guess.

point_in_time on memory_recall bounds the engram write timestamp, which is record time, not validity.

7. Write Mechanics

Writes are explicit and synchronous up to the engram: an embedding call and an off-heap write, visible to the next recall. No LLM is on the hot path. Eager contradiction checks run afterwards from a queue of capacity 256 (SpectorPropertyConstants.java:1028). Each check scans the tier and makes one LLM call per neighbour within the threshold. A contradicted flag therefore lands seconds after the write when an LLM is configured, and never otherwise. Entity and fact extraction runs in the 30-second graph enrichment job.

Agent-generated content is admitted as stated. The agent chooses the tier, the source and the id. Nothing filters content. A write under USER_STATED from an agent is indistinguishable from one a person made.

Newer wins contradictions. determineWinnerLoser compares timestamps first (CadpContradictionResolver.java:68-82). Any writer that can store a newer, contradictory SEMANTIC memory near an older one can bury the older one, subject only to the LLM's YES.

Fact extraction writes seconds. GraphEnrichmentEngine stamps extracted facts with System.currentTimeMillis() / 1000L (GraphEnrichmentEngine.java:287, :514-535). Its re-extraction path looks up a retractFactsForMemory method by reflection and swallows the NoSuchMethodException (:419-430). No such method exists in the tree, so re-extraction appends new facts beside the old ones.

Operational cost

• Write: synchronous embed and write; no LLM on the hot path.
• Eager consolidation: per semantic or procedural write, a scan of the tier and one LLM call per near neighbour; lag of seconds with an LLM configured.
• Background: graph enrichment every 30 seconds, with LLM extraction per pending memory; sleep consolidation and dreaming on their own schedules.
• Read: bounded by top_k, and by a token budget on memory_context_pack. Nothing is injected automatically, so prompt-cache placement is the caller's.

8. Agent Integration

The agent holds 29 MCP tools, among them remember, recall, graph recall, forget, purge, suppress, resolve, reinforce, fact assert, fact retract, fact history, context pack, consolidate, compile skill and persona enact. Every write and correction verb sits on the agent's own surface. The README gives setup for Claude Desktop, Cursor, Windsurf and Claude Code through npx -y @spectrayan/spector mcp. The OpenClaw plugin registers tools and skills and no lifecycle hook, so nothing injects memory at session start; the agent recalls when it chooses to.

memory_context_pack assembles a budgeted pack from recall and, for capitalised query words, fact histories (MemoryContextPackTool.java:84-93).

The default fact-history predicate is a literal. memory_fact_history defaults predicate to "*" (MemoryFactHistoryTool.java:49), and the context pack and multi-evidence recall pass "*" explicitly (MemoryContextPackTool.java:93, MemoryMultiEvidenceRecallTool.java:71). DefaultSpectorMemory.factHistory resolves it with getOrRegister (DefaultSpectorMemory.java:2234), which registers a predicate named *, and filters facts by that id. No fact carries it, so all three return no history. The tool's unit tests mock SpectorMemory with explicit predicates.

9. Reliability, Safety, and Trust

The deletion vocabulary is honest. memory_forget states that the content bytes remain on disk and in backups and names memory_purge (MemoryForgetTool.java:55-61). Purge reports bytes zeroed, graph references removed, text shared through deduplication that it could not erase, and the copies it does not reach. PurgeEndToEndTest asserts the report says "No cryptographic erasure was performed".

The fact plane's read path undoes its write path. TemporalFact.validAtInstant tests validFrom <= epochMilli && epochMilli < validTo (TemporalFact.java:136-139). The public interface documents epoch seconds (SpectorMemory.java:867-868), the MCP schema asks for seconds, the REST service defaults to seconds, and the extractor writes seconds (MemoryService.java:1280, GraphEnrichmentEngine.java:287). The kernel's own Javadoc says milliseconds (TemporalKnowledgeGraph.java:217-219). A seconds value is always below a present-day millisecond instant, so validFrom never excludes a fact and a finite validTo always does.

Retraction does not reach the reader. TemporalQuery.resolve filters retracted facts only when excludeRetracted() was called (TemporalQuery.java:208, :217), and no production code calls it. A fact retracted with no successor is still returned by factsAbout and woven into recall. The kernel test pins that behaviour: retractFactExcludesFromQueries asserts one result without the call and none with it (TemporalKnowledgeGraphTest.java:118-139). Supersession hides the old value only because LatestTxWinsResolver keeps the newest fact per predicate.

Export and import lose correction state. The importer re-asserts every exported fact row, retraction rows included, with allowCoexisting true and a fresh transaction time (SpectorMemoryImporter.java:346-370). It restores engrams through remember, dropping the exported consolidationFlags (:127-178). A round trip revives retracted facts and contradicted memories.

Human approval gates a different agent. DefaultAgentApprovalService holds every write tool of Synapse's built-in chat agent for a person at /api/v1/agent/approvals, enabled by default (DefaultAgentApprovalService.java:51-67). The queue is InMemoryAgentApprovalRepository, and the call blocks on a future for up to 300 seconds (:95-107). External MCP clients write without it.

Capability marks:

• trust_state — awarded; the contradicted flag moves by consolidation and is a filter at three read sites. Evidence and limits are in the frontmatter record.
• bitemporal — awarded on the fact log; the unit mismatch, the unused asOf and the unused excludeRetracted are its limits.
• negative_eval — awarded; evidence in section 10.
• tombstone — forget sets a bit on the record and frees the id. Suppression is an id set in process memory (SuppressionSet.java:27-37). Fact retraction is keyed on factId. Nothing keyed on a rejected value is consulted on write.
• scope_enforced — the partition is the namespace directory. No record carries a scope key and no read applies a scope predicate; the advertised workspace_id has no reader.
• audit_log — the WAL records REMEMBER, FORGET and other events and is truncated at each checkpoint (MemoryWal.java:850, CheckpointEngine.java:344). ProvenanceMemory records consolidation lineage, not mutations.
• human_review — the approval queue lives in process memory and covers only the built-in chat agent. forget, purge, suppress, consolidate and fact_retract are on the external agent's own tool surface.

10. Tests, Evals, and Benchmarks

I built and ran nothing; everything below is from reading the tests at the pin. CI runs mvn -B clean install across the reactor (.github/workflows/ci.yml:77-83), excluding the functional group (pom.xml:185), then bench property and unit tests.

The contradiction case. testContradictoryDuplicatesFlagged_cadpDirectional stores "The capital of France is Paris." and "The capital of France is Lyon." under one vector, answers the mock LLM with YES, and consolidates. It asserts the older is contradicted and the newer is not, that default recall has no fact-a and has fact-b, and that includeContradictions(true) returns both (ConsolidationIntegrationTest.java:110-156). The positive side is asserted on the same recall, so an empty result fails.

The isolation case. testCrossNamespaceRecallIsolation attaches two namespaces to one runtime, asserts each recalls its own document, then asserts neither returns the other's, including for a query built from the other's text (EngineSharingIsolationTest.java:95-141).

Purge. PurgeEndToEndTest reads raw bytes to show that forget leaves the payload intact and purge zeroes it, and checks recall, inspect and export after a purge beside an untouched survivor.

Tests named for mechanisms they do not test. GovernedMemoryE2ETest is titled for the retracted, restricted and unverified gates. Its retraction case asserts a fact and checks that recall text does not contain $50M, but no memory with that text was ever stored, so it cannot fail (GovernedMemoryE2ETest.java:44-70). The restricted and unverified cases assert only non-empty results. The whole E2E suite skips unless OLLAMA_LIVE is true (AbstractE2ETest.java:55-58).

Not covered. No test drives the fact plane through the public API with seconds-valued input and a finite valid_to; the kernel tests use milliseconds. No test exercises the "*" predicate default. No committed LongMemEval or LoCoMo result was found; the bench module carries runners and three MF-001 conformance fixtures. No paper describes the system; docs/ cites cognitive-science literature, and ADR 0064 cites one arXiv paper for a related method.

11. For Your Own Build

Steal

• Say what forget did not do. A delete verb that returns "hidden, bytes remain, use purge to destroy" stops an agent from relaying a false erasure to a user. Make the stronger verb report what it could not reach.
• Pin the corrected-value exclusion with a forensic twin. Assert the winner is present, the loser absent, and the loser back when the caller asks for contradictions. An exclusion that removes everything cannot pass all three.
• Append retractions; never mutate a fact row. The supersession chain is then a query, not a reconstruction.
• Put status gates inside the scan loop, before scoring, so a flagged record never competes for a top-k slot.

Avoid

• Two units for one field. Pick seconds or milliseconds, name it in the field (validFromMs), and make the comparison take what the writer produced. A display that guesses the unit is the symptom.
• Exclusion as an opt-in builder option. excludeRetracted as a flag means every new reader forgets it. Make exclusion the default and inclusion the flag.
• Gates on flags with no writer. A release gate with nothing to release reads as defence in depth and is none. Test it against a record the production path flagged.
• Tool schemas that promise scoping the handler ignores. The model acts on the description; an ignored workspace_id is a silent scope error.
• Newest-wins contradiction resolution without provenance. Recency is not evidence. Weigh the source, and keep the loser reachable to a person.

Fit

Spector suits a team that wants an embeddable, single-process JVM engine with fast fused recall, physical tenant separation and a large tool surface, and will run an LLM for consolidation. It does not suit a reader who needs its temporal facts to answer "what was true then": the valid-time path is broken at this commit, and retraction is invisible to readers. The governance vocabulary is ahead of the wiring, and the surface — 29 tools, many neuroscience-named relays, AISME behind a flag — is large for what is wired. Adopt it for engram recall and correction, with its contradiction test as the contract, and treat the fact plane and the release gate as unfinished.

12. Open Questions

• Do deployments assert only open-ended facts, where validFrom never binds and the unit mismatch is silent?
• Does the batch consolidator run on a schedule by default, or only through memory_consolidate?
• Which records does the eager consolidator compare when a namespace spans several partitions, given it scans one store?
• Is the NOTICE attribution clause intended as a licence condition beside Apache-2.0's section 4?
• Does the Cortex dashboard's edit surface write through the same verbs as the agent?

Appendix: File Index

• Storage and schema: memory/spector-kernel/.../engram/field/EncodingHeaderFields.java, .../engram/EncodingHeaderLayout.java, .../kernel/store/TemporalFact.java, .../kernel/store/ProvenanceMemory.java, .../kernel/scan/SlabScanner.java.
• Write path: memory/spector-memory/.../DefaultSpectorMemory.java, .../pathway/remember/RememberPathway.java, .../pathway/remember/relay/DedupGuardRelay.java, .../pathway/remember/relay/CorticalWriteTransactionRelay.java, .../graph/GraphEnrichmentEngine.java.
• Correction: .../cortex/consolidation/AbstractConsolidator.java, EagerConsolidator.java, BatchConsolidator.java, ContradictionDetector.java, CadpContradictionResolver.java; .../neuromod/inhibition/SuppressionSet.java.
• Retrieval: .../pathway/recall/RecallPathway.java, .../pathway/recall/relay/GovernedReleaseGateRelay.java, .../cortex/SemanticRecallStrategy.java, .../pathway/pipeline/gatherer/RecallCandidateGatherer.java, .../pathway/pipeline/graph/TemporalFactWeavingStage.java.
• Facts: .../graph/temporal/TemporalKnowledgeGraph.java, TemporalQuery.java, LatestTxWinsResolver.java.
• Background: .../scheduler/QuartzMemoryScheduler.java, .../sync/MemoryWal.java, .../sync/CheckpointEngine.java.
• MCP and server: synapse/spector-mcp/.../tools/SpectorToolRegistry.java, synapse/spector-mcp/.../tools/memory/*.java, synapse/spector-mcp/src/main/resources/mcp/tools/*.json, synapse/spector-synapse/.../config/McpServerConfig.java, synapse/spector-synapse/.../memory/MemoryService.java, synapse/spector-synapse/.../agent/approval/.
• Import and export: synapse/spector-batch/.../importing/SpectorMemoryImporter.java, synapse/spector-batch/.../exporting/SpectorMemoryExporter.java.
• Tests: memory/spector-memory/src/test/.../ConsolidationIntegrationTest.java, .../runtime/EngineSharingIsolationTest.java, .../PurgeEndToEndTest.java, .../graph/temporal/TemporalKnowledgeGraphTest.java, .../e2e/GovernedMemoryE2ETest.java.

Recorded searches

Checked against the checkout at the pinned revision, from the tree root.

• grep -rn --include='*.java' -E 'FLAG_RETRACTED|FLAG_UNVERIFIED|FLAG_RESTRICTED' . | grep -v '/src/test/' — only definitions, predicates and the with* setters in EncodingHeaderFields.java.
• grep -rn --include='*.java' -E 'withRetracted|withUnverified|withRestricted' . | grep -v EncodingHeaderFields.java — two test files only.
• rg -n 'writeConsolidationFlags\(|markRetracted|markUnverified|markRestricted' --type java -g '!**/src/test/**' — writes of CONTRADICTED and PURGED only, plus the cursor setter.
• rg -n '\.consolidationFlagsOverlay\(' --type java — one production setter, FLAG_CRYSTALLIZED in SkillPersistRelay.java:107.
• grep -rn --include='*.java' 'excludeRetracted()' . | grep -v '/src/test/' — the Javadoc example and the method itself.
• grep -rn --include='*.java' -E '\.asOf\(' . | grep -v '/src/test/' — no caller of TemporalQuery.asOf; the hits are graph-recall and enactment options.
• grep -rn --include='*.java' 'retractFactsForMemory' . — one hit, the reflective lookup at GraphEnrichmentEngine.java:423.
• rg -n -i 'workspace_id|workspaceId' -g '!*.lock' . — two tool schemas, ADR 0025, ADR 0062 and two Synapse tests; no production Java.
• rg -n '"\*"' --type java memory/spector-memory/src/main memory/spector-kernel/src/main synapse/spector-mcp/src/main — the three "*" defaults; no wildcard handling.
• rg -n 'truncateBefore\(' --type java -g '!**/src/test/**' — CheckpointEngine.java:344.
• rg -n -i 'class \w*(Journal|EventStore|EventLog|AuditLog|MutationLog|ChangeLog)\w*' --type java -g '!**/src/test/**' — the dream journal and the WAL journal relay; no mutation audit store.
• rg -n 'OLLAMA_LIVE|mvn ' .github/workflows/ — no OLLAMA_LIVE in any workflow.
• grep -rliE 'arxiv|bibtex|@article|@misc|doi\.org' . --exclude-dir=.git --exclude-dir=node_modules — two docs pages, ADR 0064 and a LongMemEval script; no CITATION.cff.

History

2026-09-26 — b9012992… — first reading, at the head of main, a commit dated 25 September 2026. Three marks: trust_state, bitemporal, negative_eval. Screened before reading: no auto-run surface, 2 build-time execution points (npm prepublishOnly in two packages), 9 dependency files inside the cooldown — every file in the depth-1 clone dates to the tip — and 4 unpinned surfaces; AGENTS.md was treated as data. The screen does not parse Maven. Read with grep, rg and sed; nothing installed, built or run.