Zikkaron synthesis

Type: agentic-system-analyses/types/agentic-system-synthesis.md

Bounded synthesis

Zikkaron implements persistent memory and context selection for an external conversational host, principally Claude Code. Its strongest supported adaptive contribution is automatic conversion of recorded activity into durable text and selection state that later invocations can consume. This is more than raw-log storage. It is not evidence that those changes improve future task performance. The account covers the whole shipped system at Git revision dda34a5d903d04ecb5517af214d437873c833302, with an applicability cutoff of 2026-10-15. Its evidence is static implementation, design statements and attributed outcome claims from SRC-1; no target execution, model response, completed benchmark or causal experiment was supplied. External host processing, dependencies, model artifacts and datasets are outside the frozen source. See the boundary and Source register, RT-RTE-trace-consolidation and MEM-RTE-maintenance-cycle. This synthesis uses accepted member records and makes no additional source reads.

From host request to retained state

The CLI starts an MCP service, initializes storage and retrieval, launches maintenance and attempts to install host instructions and event hooks. The service decides memory admission, transformation and selection; Claude Code still owns conversational next-action generation, tool choice and interpretation of returned text. Installed prose requests startup recall and task-outcome recording, while installed event commands can launch independent capture and retrieval scripts. Neither installation nor registration proves host execution or compliance. See RT-RTE-install-integration, RT-OBJ-zikkaron-section and RT-OBJ-hooks-config.

An explicit remember request proposes text, tags and a directory label. A predictive gate can reject it by a numerical novelty policy, with keyword/tag bypasses and a continuity discount. After admission, curation can concatenate a near duplicate, link a related item or create a row. This is computational admission, not a factual judgment or necessarily a human-approved edit. Anchors, project seeds, captured actions and derived maintenance products have different entry paths. Consequently, the claim to store only prediction errors is neither a storage-wide invariant nor a truth guarantee. See RT-RTE-remember-recall, MEM-RTE-curation-write, RT-RTE-seed-admission and RT-CLM-write-gate.

SQLite holds readable memory text alongside separate embeddings, graph relations, provenance labels and fixed-code control fields. These parts have different consequences: text supplies context, numerical representations influence similarity, and symbolic fields govern eligibility or priority. Directory and agent labels do not authenticate principals. Local model inference is wired, but exact external weight revisions are unpinned; use-dependent matrices and codebooks are distinct from pretrained models. See MEM-OBJ-memories, MEM-OBJ-selection-state, RT-CMP-embedding-engine, RT-CMP-hopfield-memory and RT-CMP-cognitive-map.

Maintenance changes what future work can receive

Post-tool capture records selected action summaries. Maintenance groups actions by directory and time, producing an activity memory only for groups of at least three; fetched smaller groups are nevertheless marked processed. Sensory episodes supply another retained input. Subsequent maintenance can abstract recurring patterns, change importance and relations, compress text, or create synthetic statements and narratives. These are automatic trace-fed writes with later consumers, not evidence of complete activity capture or successful learning. See RT-RTE-trace-consolidation, MEM-OBJ-episodes, MEM-RTE-maintenance-cycle and MEM-RTE-narrative-story.

Some checks have real force but narrow scope. Semantic promotion skips a cluster when members differ in lexical negation presence, and can reject duplicate schemas. This veto does not establish semantic consistency or truth of the generated abstraction. Memify can turn weighted co-occurrence into a stronger modified-together assertion; heat-based weight increases mean the threshold need not count actual modifications. Generated conjectures can therefore become ordinary retrievable memories without an established evidence-consuming truth-acceptance decision. This does not establish a system-complete absence of acceptance: named domain, sleep and graph internals remain partly unassessed. See EPI-RTE-cls-consistency, EPI-RTE-cls-promotion, EPI-RTE-synthetic-facts and the epistemic coverage boundary.

Recall itself can revise future memory. After selecting the current response, reconsolidation may append the query to retained text or archive old content and insert the query as a new memory. The latter branch does not disable the original row; a question is not corrective evidence. Usefulness feedback changes confidence and stability rather than independently corroborating facts. File validation checks reference freshness, sometimes returning valid when no file reference exists. These mechanisms change later reliance or selection without warranting the stored propositions. See MEM-RTE-recall-revision, EPI-RTE-feedback-update and EPI-RTE-file-validation.

Selection, delivery and recovery are separate routes

Requested recall combines configured retrieval signals, reranking, rules and context packing. Separate prompt/session hooks select retained material without a foreground recall request. Project context, hierarchical recall, story output and prospective reminders add other consumers. The external model receives selected text by value, not a reference preserving a discarded context frame. Zikkaron does not own host-window eviction. Service rules can enforce filtering at their retriever consumer, but direct SQL hooks and replay do not inherit that barrier. See RT-RTE-context-selection, RT-RTE-hook-context, MEM-RTE-project-context, MEM-RTE-hierarchy-read, MEM-RTE-trigger-delivery and RT-BAP-retrieval-rules.

Selection can introduce unsupported assumptions even without changing delivered prose. One optional ranking branch maps a camping mention to an enjoyment assertion; the static counterexample “Alice dislikes camping” demonstrates non-entailment. The inferred passage influences the original memory's rank, not necessarily its displayed content. Profile and belief rows have concrete readers, but ordinary belief-producer invocation was not established. Moreover, synthetic profile/belief dictionaries can lack heat, which the service wrapper reads unconditionally. Direct retriever/benchmark consumption remains implemented; reliable structured-result delivery through MCP recall is not established. See EPI-RTE-inferred-ranking, MEM-OBJ-derived-beliefs and the report-verification limit.

For continuity across compaction, drain updates an epoch/checkpoint or creates a placeholder; it does not acquire the host's whole conversation. Restore assembles a global active checkpoint, bounded anchors and recent memories, project material and optional predictions. Micro-checkpoints may retain only a 150-character task label. Global checkpoint replacement and global sections prevent a project-isolation guarantee. Missing hook execution or differing DB paths can split capture from consumption; swallowed errors preserve continuation rather than guarantee recovery. See RT-RTE-replay-context, RT-RTE-hook-context and RT-CLM-context-replay.

Recovery and withdrawal also depend on the route. One compressor archives originals and respects protection; independent sleep compression lacks those local safeguards. Ordinary consumers do not automatically reload archived reasons. Forget deletes the primary row and selected dependents, but not every separately derived profile, belief, trigger or narrative, nor text already delivered to a host. Thus neither universal protection nor complete semantic retraction follows. See MEM-RTE-compression-write, MEM-OBJ-memory-archives and MEM-RTE-withdrawal-path.

Evaluation and the improvement question

The representative multi-step workflow is the shipped LoCoMo evaluator, not an invented ordinary host task. An operator selects data and settings; Python owns progression. Recorded turns are automatically reformulated into retained rows, retrieval supplies selected dated text to generation, and a separate judge receives the question, gold answer and generated answer. Gold is withheld from the generation handoff. F1 and model judgments feed samples and aggregates, with optional JSON delivery. A wrong answer changes the score, not report eligibility or a repair loop. Generation/judge failures substitute values and continue; other stages lack that same recovery wrapper. See RT-RTE-evaluation-mode, EPI-RTE-answer-judgment and the workflow account.

Benchmark preparation itself is another wired trace-to-retained-context pathway: its original inputs are conversational session logs. This positive is independent of whether saved judgments later update production policy, which remains unestablished. Tool-origin updates and compaction-event checkpoint updates are additional distinct pathways. These architectural positives survive the unresolved mixed-origin and library branches; they do not establish observed activation or benefit. See RT-RTE-trace-consolidation, RT-RTE-replay-context, RT-RTE-evaluation-mode and the profile verification.

Improved capacity for future action remains unestablished at the memory-engine boundary. For the claimed memify self-improvement pathway, the horizon is maintenance cycles followed by later retrieval, and the independent proposed objective is useful future retrieval. State changes and later consumers are wired. Evidence that objective-bearing outcomes determine these updates, followed by subsequent operation causally depending on them, is not established. The four-link exercised self-improvement attribution therefore remains open; the source's label does not close it. No criticism-of-consumed-theory-to-capacity-gain chain is established either. The accepted runtime does not supply a formulated-theory assessment sufficient to assign theory-builder conditions 1–4 separately; none is inferred from storage, rule execution, scores or repeated cycles. Reflection and autonomous theory building likewise remain unestablished as system properties, independently of the code-owned automation already described. See EPI-CLM-derived-facts, EPI-RTE-feedback-update and the bounded epistemic conclusion.

For a persistent-context scenario, the supported result is plural, adaptive selection of retained material, not exact conversation replay or globally enforced project-local knowledge. For a QA measurement scenario, the supported result is a concrete retrieval/generation/scoring pipeline, not reproduced performance or a component-effect estimate. README outcome claims remain attributed reports, with a disclosed reader-model difference confounding the QA contrast. Host traces linking capture, update and later consumption would establish operation; controlled comparisons holding the reader and data fixed could test capacity gain. Evidence-consuming acceptance records could change the knowledge assessment. Scoped checkpoint selection, shared DB configuration, structured-result normalization and consistent protection/withdrawal handling would change the corresponding architectural limits, but would still require execution evidence for reliability claims. See EPI-CLM-benchmark-benefit, RT-RTE-replay-context and report verification.

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
No observed operation or causal outcome comparison	SRC-1; RT-RTE-trace-consolidation, RT-RTE-evaluation-mode, EPI-CLM-benchmark-benefit	Static shipped implementation and attributed README claims at the pin; no datasets, responses or result artifacts	Successful deployment, activation, reproduced scores, improved capacity and component-specific benefit	Inspectable traces and outcome artifacts; controlled interventions with fixed readers, data and model identities
Structured-result service incompatibility carried from report verification	SRC-1; MEM-OBJ-user-profiles, MEM-OBJ-derived-beliefs, EPI-RTE-profile-beliefs, RT-RTE-context-selection; verification limit	Synthetic retriever dictionaries omit heat; MCP wrapper reads it unconditionally	Reliable or unchanged profile/belief delivery through ordinary MCP recall; actual exception incidence	A normalized result contract and executions covering surviving structured results; direct benchmark consumption remains implemented
External host and model contracts remain unfrozen	SRC-1; RT-RTE-install-integration, RT-RTE-hook-context, RT-CMP-embedding-engine, RT-CMP-evaluation-models	Installed files and callers, not actual grants, host processing, dependency internals or exact external artifacts	Compliance, deployment isolation, fixed numerical behavior and hard latency guarantees	Host configuration/execution evidence, resolved artifacts and dependency identities, measured timing across fallback paths
Partly inspected transformation and consumer alternatives	SRC-1; MEM-RTE-maintenance-cycle, MEM-RTE-hierarchy-read, EPI-RTE-causal-outputs, MEM-RTE-sync-import	Full domain/community/PC/conflict-resolution semantics and some resource or external-library callers unresolved	System-complete negatives about acceptance, complete curation/read-back enumeration, causal identification and comprehensive semantic correctness	Remaining algorithms, their assumptions, production callers and concrete checked outputs
Initial provenance and admission control are unresolved on caller/library branches	SRC-1; RT-RTE-remember-recall, MEM-OBJ-derived-beliefs, MEM-RTE-sync-import, MEM-OBJ-index-enrichment	Known subsequent compilation and readers do not determine initial origins, human decisions or deployed library wiring	Complete lineage, universal manual/automatic classification and authenticated source warrant	Origin/admission records, caller identity/control evidence and deployed producer traces
Enrichment representations can disagree	SRC-1; MEM-OBJ-index-enrichment, EPI-RTE-index-enrichment	Row embedding replacement versus original-variable vector insertion	Uniform enriched-vector influence or equivalent representations across consumers	Explicitly aligned updates and consumer-specific retrieval traces
Continuity, protection and withdrawal are path-specific	SRC-1; RT-RTE-replay-context, MEM-RTE-compression-write, MEM-RTE-withdrawal-path	Global checkpoints and bounded sections; different compression and deletion bodies; host window external	Exact replay, project isolation, universal protection, rollback or complete derivative retraction	Full capture evidence, scoped replay tests, consistent archive/protection enforcement and dependency-aware withdrawal traces
No closed theory-criticism or exercised improvement attribution	SRC-1; EPI-CLM-derived-facts, EPI-RTE-answer-judgment, EPI-RTE-feedback-update	Accepted records show automatic updates and measurement, not a complete formulated-theory account or objective-responsive update-to-later-operation chain	Theory-builder conditions 1–4, criticism-caused capacity gain, reflective/autonomous theory building and exercised self-improvement	Separately grounded theory/criticism/iteration records; objective-bearing evidence selecting internal changes and later operation demonstrably depending on them