MGI-Mind synthesis

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

Bounded synthesis

MGI-Mind is a durable memory system for assistants, not their enclosing problem-solving runtime. Its distinguishing contribution is to separate retained content from the state that controls its admission, visibility and priority. Notes, fact triples, error/fix procedures, named skills and session summaries have different consumers. Reported outcomes and query/access activity can change durable selection state without changing model weights or rewriting a playbook. Those mechanisms are implemented; successful host reliance and improved future capacity are not demonstrated.

This account uses code-grounded records for the shipped system at Git commit a8defc36f7d0bcd5108714763a12625a0eda0868, with applicability cutoff 2026-11-08. SRC-1 is the repository https://github.com/madgodinc/mgi-mind. The whole-system functional boundary includes the Rust binary, CLI/MCP/HTTP surfaces, maintenance, shipped Python clients and integration drivers, prompts and evaluators. It excludes third-party host/provider internals and live deployments. No target commands, tests or model services were executed, and no inspected run or causal experiment establishes efficacy. “Whole-system” identifies the shipped responsibilities, not exhaustive inspection of every implementation branch; an additional opaque HTTP persistence surface remains outside the characterized memory inventory. See the boundary and evidence, Source register and report-verification limit. This synthesis adds no source reads.

From assistant request to retained material

An assistant or operator decides when to use memory. The binary offers direct CLI commands, stdio MCP and HTTP service modes. MCP dispatch is serial JSON-RPC; HTTP uses separate handlers and access controls. The shipped OpenAI Agents driver exposes only remember and recall, while broader MCP registration can expose other stores and maintenance operations. Installing tools does not install a habit: the operating manual requires a separate pointer in the host's session-loaded rules. The integration's instruction to recall before dependent answers is natural-language guidance, not a mandatory server gate (RT-RTE-assistant-memory, RT-CLM-recall-habit, boundary).

The representative workflow asks the assistant to remember a production database address, then asks its port. The model must choose to call each tool. Remember sends caller text through the Python HTTP client to direct storage; storage checks library registration and secrets, chunks and embeds it, and writes Qdrant points. Recall searches the configured library, returns text through a tool result, and the external host generates an answer. Delivery means printing the host's final_output. The driver neither asserts the expected port nor verifies that either memory call occurred. It shows no transfer of the first conversation into the second Runner call and no local retry or repair handler; SDK continuation and exception recovery remain external (workflow account, WF-RTE-response-delivery). This is a conditional implemented workflow, not an observed successful remember/answer episode.

Admission depends on the route. Direct add, web save and Markdown reconciliation do not pass ingest's additional relevance/novelty policy. Ingest can accept, skip or quarantine a candidate, and failed neighbor lookup can accept novelty by default. Quarantine preserves content for later review or promotion rather than proving it false. Transcript harvesting selects role-tagged user/assistant text and excludes tool/thinking blocks; it discards its computed session-specific source tag. Provenance import checks citation shape without fetching or authenticating the cited passage. These controls concern storage suitability and lineage, not evidence-consuming acceptance of proposition truth (RT-RTE-ingest-admission, MEM-OBJ-quarantine, MEM-RTE-transcript-harvest, EPI-RTE-provenance-import). Software admission does not require a per-item human decision; initiating an import or selecting apply is not evidence of such approval.

The optional extractor adds a further transformation. In an extractor-enabled build with the server installed, newly admitted memory can enqueue a bounded background job. A local model proposes triples; JSON parsing and one wording repair enforce response structure before global fact admission. Queue overflow and failures can drop work, and the original caller does not receive a guaranteed extraction result. This supports conditional standing-fact consolidation, including from harvested session messages, without establishing actual output fidelity. Without paired inputs/outputs and checked interpretation, whether a particular triple merely reshapes, derives or conjectures remains indeterminate. Parsing is not semantic verification (RT-RTE-auto-extraction, EPI-RTE-extraction-parser, MEM-RTE-fact-lifecycle).

What persists and what a later consumer receives

Notes retain prose and metadata in Qdrant, alongside dense and sparse retrieval indexes. Facts are vectorless triples with separate lifecycle/cardinality state. Procedures store error/context/fix text; skills store named trigger/body playbooks. Session files and pinned blocks are local files, while procedure/skill outcomes occupy separate derived statistics. These are not interchangeable representations of one memory: numerical indexes select content, symbolic flags constrain server behavior, and a host interprets returned prose (MEM-OBJ-memory, MEM-OBJ-memory-vectors, MEM-OBJ-fact, MEM-OBJ-fact-state, MEM-OBJ-procedure, MEM-OBJ-skill, MEM-OBJ-procstats).

Ordinary host recall is requested, not unsolicited context injection. Note search fuses filtered dense and lexical results, optionally reranks full content, then limits and display-truncates it. Rerank failure retains the fused order. The representative wrapper uses tier two: at most 500 characters per returned hit. Python renders content and optional author, omitting structured hit IDs and scores. Thus ranking has access to more text than the assistant receives, and neither retrieval nor its displayed order establishes truth (RT-RTE-assistant-memory, WF-RTE-response-delivery).

Other requested reads differ. Unified recall combines facts, notes and procedures, not skills, with per-silo limits rather than a joint score or global token budget; an omitted section can reflect a swallowed silo error. Context returns a briefing containing pinned blocks, a skill catalogue, prior-session material and selected facts. It uses the globally newest session, not an agent-specific session. It selects up to 20 newest valid/default-visible fact candidates before checking only valid_until; it does not check stored valid_from. Skill bodies are kept out in favor of names/triggers, with match/show as reload routes. This is selective supply by value with a reloadable catalogue, not demonstrated host compaction or an isolated child-agent frame. The host still owns injection, accumulation and eviction (MEM-RTE-unified-recall, RT-RTE-context-briefing).

Revision, feedback and the authority they actually have

Fact handling changes operational standing. On declared single-valued axes, deterministic comparisons use incumbent history and challenger weight to select active, contested or quarantined standing and retire displaced entries. First writes and Multi axes bypass duels. Contested and unknown facts can remain default-visible. Explicit invalidation vetoes ordinary reassertion until explicit restore; restoration does not automatically reverse a stale lifecycle status. These are effective server controls, not comparisons with world truth (EPI-RTE-fact-comparison, EPI-RTE-fact-disposition, EPI-RTE-fact-withdrawal, EPI-RTE-fact-consumption).

The doubt mechanism compares retained origin/query vectors, not assertions against new observations. Shadow mode measures without changing ranking; enforce mode requires the relevant vectors and can update counters and reorder facts. Background retest automatically supplies bounded fact candidates to a confidence formula. This is a server-side push to an evaluator, distinct from requested host recall. Retest recomputes metadata confidence rather than empirically testing the proposition. Moreover, typed-signal helpers address the memories collection while fact scalars reside in the facts collection; the generic typed-outcome-to-fact feed is not established. That discontinuity does not erase scalar retest or the working procedure bridge (MEM-RTE-fact-retest, EPI-RTE-drift-evaluation, EPI-RTE-confidence-retest, EPI-CLM-outcome-duel).

The strongest supported adaptive contribution is durable outcome-responsive ranking. Reported procedure/skill reuse updates counters consumed by later recall/match. A new typed test/compile signal can bridge to procedure statistics; a verified flag adds a ranking boost, and outcome ratios can demote failures. But labels and caller flags do not authenticate test execution. Verification latches true after later failures, and replacing a same-key signal can change its log without reversing the earlier counter effect. This is trace-fed behavior-shaping state, not established improved capacity or content-directed criticism. Query-derived drift state, apply-gated access-based coldness maintenance and conditional transcript-to-fact extraction supply additional trace-fed paths. Their original inputs and activation conditions remain distinct (MEM-RTE-outcome-ranking, EPI-RTE-outcome-trust, MEM-OBJ-access, RT-RTE-store-consolidation, profile verification).

The retained records do not establish theory-builder conditions 1–4 for actual playbooks: condition 1 needs individuated theory content, condition 2 decisions depending on that content, condition 3 stated content-directed criticism with changed reliance or revision, and condition 4 retention shaping another round. A mutable skill API and outcome counters alone establish none of that sequence. Nor is criticism of a consumed theory shown to improve future capacity. Reflective and autonomous theory-building properties remain independently unestablished; computation owning server admission/ranking is not autonomous ownership of every theory-building role. Exercised self-improvement is also unestablished: durable ranking changes have a live later-consumer path in code, but no subsequent operation is shown causally depending on such a change, and no outcome comparison establishes success. These are limits of the assessed routes, not assertions that opaque models or unassessed branches lack those properties (RT-RTE-skill-admission, MEM-OBJ-skill, EPI-RTE-outcome-trust, epistemic conclusion).

Recovery is likewise local. Quarantine promotion, archive restoration and explicit fact restoration differ from deletion. Markdown replacement deletes prior points before rewriting and is not an atomic rollback. Skill replacement can retain trust history without verifying the new body. Portable exports render verified fixes for possible host-file consumption, but neither install nor revoke saved copies automatically (RT-RTE-markdown-reconcile, RT-RTE-skill-admission, MEM-RTE-portable-export).

Scenario-relative assessment and evidence that would change it

For cross-session retrieval of an assistant-supplied fact, the shipped driver establishes the necessary conditional write/search/return chain, not reliable use or answer correctness. For avoiding stale or contradictory assertions, MGI-Mind implements explicit withdrawal, conflict standing and conditional context-sensitive ranking, not a general truth-maintenance guarantee. For improving reuse of fixes, it implements reported-outcome ranking, not authenticated success, criticism of explanations or demonstrated transfer. Its evaluators measure gold session IDs, exact fix strings or authored scenario expectations. Procedural benchmarking does not exercise trust-ranked recall and its temporary-library cleanup does not isolate actual procedure writes. The policy baseline is constructed, and frozen STALE dispatch aggregates no scenarios. No published score is adopted here (EPI-RTE-benchmark-evaluation, MEM-CLM-benchmark-isolation).

Consequential changes would be inspectable host traces connecting admitted content to later tool use and answers; version-linked authenticated outcome evidence; repaired fact-signal routing; and independently varied comparisons of ranking or validity controls. Complete interval filtering would require checking the stored lower bound in briefing. Universal scoped-write exclusion would require propagating confinement through accepted-memory extraction, not merely its explicit candidate branches. Today HTTP normalizes library filters and rejects scoped text recall, but a scoped admitted Memory candidate can enqueue global fact extraction when feature/install/admission conditions hold. Actual triples and successful global writes remain unobserved. CLI, trusted stdio and local storage access are separate trust routes, not governed by HTTP tokens (RT-RTE-scoped-http, RT-RTE-auto-extraction).

Limitations

Limitation Affected SRC IDs and record citations Inspected boundary Conclusion prevented Evidence that would resolve it
Additional opaque JSON KV persistence is not characterized in the analyst inventory. SRC-1; report-verification limit Included HTTP /kv/set and /kv/get persist arbitrary JSON to data_dir/kv_store.json; no later assistant consumer established. Complete storage inventory or comparison of all checkpoint/resume channels. Established note/fact/playbook/session routes remain supported. Records characterizing KV payloads, admission and actual later consumers, or a bounded absence of such consumers.
No observed host workflow or causal efficacy evidence is established. SRC-1; RT-RTE-assistant-memory, WF-RTE-response-delivery, EPI-RTE-outcome-trust Shipped drivers, client and server wiring; external SDK/provider internals and deployments uninspected. Reliable tool use, correct answers, improved capacity, exercised self-improvement or component benefit. Traces linking versions, writes, returns and later decisions, plus comparisons suited to the claimed effect.
Extraction semantics and activation are unobserved. SRC-1; RT-RTE-auto-extraction, EPI-RTE-extraction-parser Default-off feature, installed extractor and admitted-memory branch; parser checks structure. Faithful extraction, entailed output, produced accepted ampliative claims or observed global-write effects. Paired inputs/outputs, semantic evaluation and run evidence through fact admission and later consumption.
Fact typed-signal continuity is not established. SRC-1; MEM-RTE-fact-retest, EPI-CLM-outcome-duel Fact scalars use facts collection; typed helpers use memories; challenger has no typed score. Generic outcome-fed fact duels or fresh empirical re-justification of facts. A supported fact-local producer/bridge and consequential consumer, with linked evidence.
Outcome standing is caller-reported and may outlive corrections or body changes. SRC-1; MEM-RTE-outcome-ranking, RT-RTE-skill-admission Typed novelty bridge, verified latch, stable skill identities and ranking consumers. Authenticated testing or current correctness from verified status. Version-linked test provenance and correction/revocation semantics exercised through ranking/export.
Briefing is neither complete temporal filtering nor host context control. SRC-1; RT-RTE-context-briefing, MEM-RTE-unified-recall Global newest session, upper-bound-only fact check, per-surface limits, catalogue reload. Agent-specific continuity, full stored-interval validity, complete recall, global token bounds or host compaction. Lower-bound-aware selection and host traces showing actual selection, injection and eviction.
Maintenance, initialization and consumer coverage remains partial. SRC-1; memory inventory, MEM-OBJ-origin-context, MEM-OBJ-session, MEM-OBJ-vault, profile verification Migration/backfill, cache initialization, complete restore/reassembly, history/as-of semantics, retest priority, summary production, arbitrary initial content and opaque vault values are incompletely resolved. Complete memory-value inventories, uniform lineage, exact temporal/restore equivalence or exhaustive safety claims. Accepted records tracing each unresolved part, transformation and consumer at the frozen boundary.
Evaluation supports bounded metrics, not efficacy attribution. SRC-1; EPI-RTE-benchmark-evaluation, MEM-CLM-benchmark-isolation Evaluator code; raw retained results uninspected; procedural isolation mismatch, constructed policy baseline and empty STALE aggregation. Clean-room procedural results, fix execution success, answer quality, transfer or causal ranking benefit. Isolated evaluator writes, authenticated run artifacts, real intervention baselines and appropriate outcome measures.
Actual theory content and criticism are unresolved independently of trace adaptation. SRC-1; MEM-OBJ-skill, RT-RTE-skill-admission, EPI-RTE-outcome-trust Mutable body admission and retrieval plus content-insensitive counters, without actual playbook episodes. Theory-builder conditions, reflective/autonomous theory building or criticism-driven learning. Individuated content, consumption, stated criticism and retained next-round consequences, with separate evidence for each property.
HTTP controls are route-specific, not universal confinement. SRC-1; RT-RTE-scoped-http, RT-RTE-auto-extraction Scoped filters/text rejection and conditional accepted-memory enqueue; separate CLI/stdio/local routes. Universal global-write exclusion or deployed isolation; no successful bypass is demonstrated. Scope propagation through all side effects and deployment evidence covering storage/filesystem access.