RightMemory synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Retain context without taking over the coding task
RightMemory is a memory and context-engineering system for coding agents, not the enclosing coding runtime. This account covers its shipped lifecycle at commit 62534bd82c58718f31c99692fed4c5d159e218ec, with an applicability cutoff of 2026-11-08. The whole-system source boundary includes internal model execution, native-host instructions, maintenance, synchronization, sharing and evaluators. SRC-1 supplies implementation, design instructions and attributed reports; no observed-run or causal-experiment evidence is registered. External hosts, model services, deployed memory roots and private benchmark inputs are outside observation. See the boundary and evidence and Source register.
The system separates three semantic responsibilities. Memory stores durable facts, decisions, preferences, guidance and open questions. Pursuit stores user-owned directions in a related document graph; its Focus marks attention, not an execution queue or permission to act. Agent Corrections stores reusable contrasts between prior agent work, user redirection and a settled result. These correction cases are fixed non-graph collections, not ordinary graph nodes. Markdown carries interpreted meaning; IDs, edges and ranges have fixed selection effects. Linked evidence documents and complete instruction assets are distinct from graph continuation. See MEM-OBJ-memory, MEM-OBJ-pursuits, MEM-OBJ-agent-corrections, MEM-OBJ-markdown-evidence and MEM-OBJ-memory-skill.
Select context, then let the host interpret it
A caller requests context through one of three materially different organizations:
- Internal agent retrieval: RightMemory supplies a root snapshot, subsequent differences, pending evidence and delivery-change information to a role model. The model selects source IDs, ranges or cases; program code resolves those selections into source-derived text. Successful terminal output is not the model's free-form account. Invalid selections receive applicable bounded retries; oversized output is rejected rather than truncated. This protects source resolution and output form, not relevance or completeness. The final character limit also does not bound model input: current versions of previously returned content use a separate effectively unbounded renderer. See RT-RTE-agent-retrieval and its memory annotation, and MEM-OBJ-retrieve-context.
- Embedding retrieval: RightMemory indexes graph text, evidence, complete instruction assets, imported content, correction cases and pending submissions. It ranks entries by their best passage similarity, reranks a candidate set, applies count limits and renders original source text in source order. A corpus-fingerprint check retries a changed corpus once. Queries may return prior results; internal agent delivery-history semantics do not apply. Only embedding-service errors trigger configured agent fallback; oversized-render errors do not. Fixed-count selection is not answer-absence detection. See RT-RTE-embedding-retrieval and its memory annotation, and MEM-OBJ-vectors.
- Native retrieval: A model-free service prepares instructions, pending evidence and live source paths. A host-native child reads files and authors the context handoff. It does not inherit the internal selector, output budget or delivery ledger. The host owns child scheduling and context isolation; the reported commit does not freeze subsequent live-file reads. See RT-RTE-native-assistant and the workflow account.
Requested delivery is distinct from automatic supply inside an internal role invocation: new-conversation snapshots, broad change/pending notices and identity-targeted notices about previously delivered sources can enter that invocation without a separate request for each part. Saved histories support continuity, while host/provider compaction remains unresolved. Neither separate child contexts nor pruning establishes general management of the enclosing host's token window. Retrieved prose supplies context or guidance through interpretation; the shipped skills request actual application subject to current evidence and user instructions, but compliance is unobserved. See MEM-OBJ-retrieve-context, RT-BAP-host-context and RT-RTE-maintenance-cycles.
Submit evidence, judge meaning, admit committed state
An update submission is pending evidence, not final stored wording or settled truth. It can be returned by retrieval while still pending. Update receives submitted evidence and current state, forms keep/merge/skip/edit judgments, and then consults separate edit-feedback examples as a late check. Those examples preserve candidate/proposed/accepted edit contrasts and are excluded from ordinary retrieval context. They criticize curation, whereas Agent Corrections assists future task work. Immutable batches retain inputs without guaranteeing that every resulting assertion carries a derivation link or that every completed batch has a later semantic reader. See MEM-OBJ-update-evidence, MEM-OBJ-edit-feedback and RT-RTE-isolated-update.
The model's semantic judgment and the program's publication decision are separate. Update is asked to assess durability, future usefulness, scope, duplication and evidence clarity, revalidate stale claims and retain consequential uncertainty rather than invent facts. The supervisor can reject dirty or uncommitted work, forbidden paths, invalid structures and stale semantic bases; intermediate commits are checked before collapse. These checks establish structural admissibility, not factual correctness or accepted novel knowledge. Correction-capacity enforcement is deliberately disabled in whole-memory validation, and Insight skips that graph-validation stage while retaining path/tree checks. Ordinary root configuration invokes isolation; shipped nested execution has a direct branch because it is already operating inside the supervisor's worktree. That branch is not evidence of a public CLI bypass. Provider-native write tools can have broad capabilities before publication, so commit checks are not an OS sandbox or universal external-effect rollback. See EPI-CLM-memory-admission, EPI-CLM-correction-cases, RT-RTE-isolated-update and RT-RTE-maintenance-cycles.
The representative native workflow connects these steps explicitly. The main host delegates a context need, performs the user's work, submits consequential evidence under stable session/request identity, and delegates a newly claimed Update batch to one native child. Repeating the same identity with different evidence is rejected. An existing task is not permission to launch a duplicate; lease expiry alone does not abandon edits. The child commits in a persistent isolated worktree, and program completion validates/imports those commits and produces a landed-state receipt. Stale edits can be corrected in the same worktree; uncertain completion calls for status inspection and same-task completion retry, not immediate replacement. Prepared work is not publication, and the memory receipt does not certify delivery of the user's coding task. This is a source-supported protocol, not a demonstrated successful execution. See the workflow account, RT-RTE-native-assistant and RT-CLM-native-publication.
Maintain and share through different authorities
Update is not the universal admission route. Explicit guidance review requires user approval of proposed meaning and scope before host curation, although the procedure is instructional rather than a mandatory software gate. The human-facing Pursuit editor validates allowed map changes and journals undo/redo state; requested opening context projects one Pursuit's ancestors and incoming/outgoing one-hop neighbors, not arbitrary transitive context. Pruner uses generation, semantic-use and grace criteria to reduce active memory, not a truth test or FIFO expiry. Insight writes separate logs that later Insight cycles may read, without automatic promotion to active Memory. Direct maintenance and synchronization have their own authority and checks. See MEM-RTE-guidance-review, RT-RTE-pursuit-editor, MEM-RTE-opening-context, RT-RTE-maintenance-cycles, RT-RTE-manual-maintenance and RT-RTE-sync-admission.
Shared file views can be extracted or generated. Extractive rendering preserves surviving source wording, not necessarily all context needed to interpret it. Generative rewriting has no established meaning-equivalence guarantee. Approval licenses distribution rather than accepting propositions as true; semantic refresh preserves old approval instead of requiring renewed approval of every output. HTTP import checks package identity, namespace and structure before promotion; failed refresh can retain a valid stale import. A separate, explicitly requested question view invokes provider policy and returns an answered/unavailable response, not a truth endorsement. Thus the internal source-rendering guarantee cannot be generalized to every shipped answer path, and imported instructions are not automatically installed or executed. See EPI-RTE-view-build, EPI-RTE-view-approval, MEM-RTE-view-refresh, RT-RTE-file-import, MEM-RTE-question-view and MEM-OBJ-shared-view.
What changes across later work—and what remains unestablished
The strongest supported learning contribution is an automatic route from submitted conversational work and user redirection to durable guidance or correction cases, then to a future host through retrieval. This is a wired trace-fed update capacity, not merely raw-log retention and not model-weight training. It does not establish improved capacity for future action, nor that criticism of a consumed theory produced such improvement. The unresolved comparison is between the same later work with and without the retained change; publication and delivery alone supply no outcome or causal evidence. See MEM-OBJ-update-evidence, MEM-OBJ-agent-corrections, RT-RTE-isolated-update, RT-BAP-host-context and profile verification.
The reports do not establish theory-builder conditions 1–4 as a completed system property. They identify localized guidance and interpreted consumption paths, but do not separately establish a formulated theory, decisions dependent on its content, content-directed criticism with blame and changed reliance, and retention of that criticism shaping another round. Neither correction cases nor Insight's name closes those gaps. Reflection is independently unestablished: prior-log use is not evidence of both causal directions between a selected system aspect and its self-representation. Autonomous theory building is also unestablished; model-owned semantic decisions do not demonstrate computation performing every internal theory-building operation. Exercised self-improvement at this shipped memory-system boundary remains unestablished: retained guidance can enter later behavioral-authority paths, but subsequent operation causally dependent on an objective-bearing update is not supplied. These are limits on attribution, not evidenced absences or a grade for the system. See the epistemic bounded conclusion, RT-RTE-maintenance-cycles and RT-BAP-host-context.
For preserving selected source text and admitting structurally bounded changes, the inspected architecture supplies concrete controls. For deciding relevance, factual warrant, privacy adequacy of generated shares or improved coding outcomes, it supplies fallible semantic policies or incomplete evidence. The repository's reported 28/28 positive required-ID cases concern a small agent-labelled retrieval test, not independently evaluated answer quality; its no-answer queries still receive candidates. Input/output traces could establish which semantic transformations actually occurred; later-consumer traces could establish activation; controlled comparisons could test downstream benefit without assigning component effects more finely than the intervention. A per-output semantic approval gate or demonstrated remote withdrawal mechanism would change the present sharing assessment. See EPI-CLM-retrieval-coverage, EPI-RTE-view-build, MEM-RTE-view-refresh and report verification.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No registered execution or causal comparison | SRC-1; RT-RTE-isolated-update, RT-BAP-host-context, RT-RTE-native-assistant | Shipped implementation and instructions, not deployed work | Successful semantic compliance, activation, improved capacity and exercised self-improvement | Linked update, publication, later-consumption and behavior traces; outcome comparisons at the claimed attribution level |
| Theory and reflection assessments lack a complete supported route | SRC-1; RT-RTE-maintenance-cycles; epistemic conclusion | Named Update, correction and Insight paths; no system-complete absence search | Conditions 1–4, reflective or autonomous theory building, or a system-wide negative | Records of formulated theory, consequential consumption, explicit criticism/blame, retained next-round effects and role ownership; separate self-representation evidence |
| Coarse behavioral-authority classification groups distinct consumers | SRC-1; MEM-OBJ-memory, MEM-OBJ-insight-logs, MEM-OBJ-vectors, RT-OBJ-semantic-operation, MEM-RTE-question-view; profile verification limit | Supported existence findings and partial coverage, not a complete part-by-consumer matrix | Per-consumer inventory counts or inference that every grouped part has every listed force | Finer part/consumer units preserving each actual channel, force and horizon |
| Deployed lineage and initial admission remain unresolved | SRC-1; MEM-OBJ-memory, MEM-OBJ-edit-feedback, MEM-OBJ-update-evidence | Seeds, subsequent update routes and late feedback consumption | Complete trace provenance, initial human approval or authentic source warrant for real roots | Deployment initialization records, original evidence and claim-level lineage/admission records |
| External model and continuation state is not fixed or fully inspectable | SRC-1; RT-CMP-role-model, RT-CMP-embedding-model, RT-CMP-reranker-model, MEM-OBJ-retrieve-context | Caller/adapters and readable local state, not provider internals | Fixed-model reproducibility, complete continuation classification or general host compaction guarantees | Exact deployed model identities, provider state contracts and inspectable continuation/window traces |
| Some transport, withdrawal and recovery alternatives are not fully audited | SRC-1; RT-RTE-file-import, RT-RTE-sync-admission, MEM-OBJ-shared-view, RT-RTE-maintenance-cycles | HTTP candidate promotion, Git-file dispatch and named recovery paths; not all hub/transport/history combinations | Universal transport validation, immediate purge of delivered content or complete rollback/restoration | Full relevant branch traces and remote-consumer withdrawal/recovery evidence |
| Generative meaning and approval principal are unresolved | SRC-1; EPI-RTE-view-build, EPI-RTE-view-approval, MEM-RTE-view-refresh | Extractive implementation and generative/approval wrappers | Semantic fidelity, privacy adequacy, newly accepted propositions or per-output human approval | Actual source/output pairs, semantic checks and identified approval decisions covering refreshed output |
| Additional stored evidence and feedback do not establish another automatic consumer | SRC-1; MEM-OBJ-update-evidence, MEM-OBJ-shared-view | Completed-batch storage and provider-note/list capabilities | Universal completed-batch rereading or autonomous provider-note adoption | Traced later consumers and their admission/update effects |
| Benchmark support is reported and narrowly scoped | SRC-1; EPI-CLM-retrieval-coverage | Tracked experiment report; private corpus, labels and raw results unavailable | Independent coverage/latency verification, factual-answer accuracy or downstream causal benefit | Frozen inputs, labels, raw results and suitable independent outcome comparisons |
No source files were read for this synthesis; its findings use the supplied accepted members. The report verification declares no limits or blockers; the profile verification's coarse-unit limit is carried above. No unresolved cross-report conflict was supplied.