Pond synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From external sessions to a durable archive
Pond is an archive and recall system for agent sessions, not the runtime that decides an agent's next reasoning step. This account covers the shipped Rust application, adapters, prompts, Pi/OpenClaw/Hermes integrations, maintenance, restoration consumers and evaluation material at commit 0c440e130faf44052d0d896a585e572090079f1f, with an applicability cutoff of 2026-10-14. SRC-1 supplies inspected implementation, declared contracts and attributed measurements. No target executable, test, provider or deployment was run; wired paths are not observed successful operation. The whole-system boundary includes material shipped callers, but does not imply exhaustive inspection of every codec or dependency. See boundary and evidence and Source register.
External harnesses produce native logs. Configured adapters turn them into canonical sessions, messages and ordered parts; HTTP canonical ingest is another entry. Pond's software validates structure, strips caller-supplied core provenance, filters existing keys and preserves stored immutable labels. Human authorship of a message and starting sync are not per-claim approval. Managed Pi serving can bootstrap capture only when its configuration preconditions hold, then sync periodically. A recorded capture decline suppresses its bootstrap argument. These choices govern future capture, not access to already stored history: RT-RTE-adapter-config, RT-RTE-session-capture, EPI-RTE-ingest-disposition.
The archive retains identity and lineage metadata, conversational text, reasoning, tool arguments/results, files and codec-specific source payloads. These parts have different readers. Full-message, export and restoration readers can load complete canonical parts, while summary readers need not load file blobs. Retention is therefore broader than the ordinary searchable view. It is also bounded: shared extraction caps individual string leaves at 10 MiB, including strings inside retained raw JSON. “Lossless” preservation is not an unconditional guarantee: MEM-OBJ-session, MEM-OBJ-part, MEM-OBJ-raw-record, EPI-RTE-field-extraction.
Messages, parts and session rows commit separately. Failures can leave durable partial progress; available source records and a successful later retry can complete missing rows. This is additive recovery, not whole-session atomicity. Bulk copy/archive import and library writes are distinct admission paths, so wire-validation guarantees do not cover every write. Copy verification checks key coverage, duplicates and index conditions, not payload equality: RT-CLM-ingest-recovery, MEM-RTE-archive-transfer, EPI-RTE-copy-verification.
From stored history to selected context
Pond derives search_text from conversational User/Assistant text and selected file descriptors. It excludes tool bodies, reasoning, injected scaffolding, System carriers and file bytes. Full-text search uses this projection; optional semantic search embeds the same restricted text with a 512-token inference bound. Message vectors persist and affect later ranking; the loaded embedding weights are a separate component and are not trained by this path. Their upstream model ID is mutable rather than an exact artifact pin: MEM-OBJ-search-text, MEM-RTE-message-embedding, RT-CMP-candle-embedder.
The host model or user chooses queries, filters and follow-up IDs. Pond selects one retrieval arm, renders results and returns copied text. Vector requests are refused when embedding is disabled; enabled instances without vectors fall back to full-text search. Pi defaults to ten search hits, clamps to twenty-five and clips response text at 32 KiB before adding a notice. Session reads are paged; message reads can expand parts and bounded neighbors. These are result and byte bounds, not tokenizer guarantees or Pond-controlled host compaction. Omitted content remains reachable through another explicit request: RT-RTE-archive-recall and its memory annotations, EPI-RTE-recall-rendering.
A representative shipped workflow answers a question about a prior decision: the host searches, selects a hit, reads the session's end for latest-state evidence, optionally pages or expands a message, then answers. Pond supplies connected tools and instructions, but neither automatically chains these calls nor dispatches a reasoning worker. Navigation errors and disabled-vector refusals return diagnostics; the caller owns correction and retry. Delivery by Pond is a tool result, not acceptance or delivery of a final answer by the host. The workflow is wired and instructed, not an observed completed episode: workflow account, RT-BAP-recall-guidance.
For this scenario, the central strength is recoverable historical evidence with bounded first reads. The central risk is mistaking selected history for complete or current evidence. Search can retrieve a confident early statement that later turns overturn. The shipped instruction to read the end is advisory, not a semantic withdrawal mechanism. Its particular SQL fallback over search_text cannot recover excluded tool output either; message expansion or a body-targeted SQL query over parts.variant_data is needed. Neither establishes that the archive covers every past event: MEM-BAP-recalled-history, MEM-OBJ-search-text, workflow SQL annotation, RT-RTE-corpus-sql.
Scope controls must also be read locally. OpenClaw filters search and withholds factory-projected tools from sub-agents, but does not independently verify the project of each targeted get. Pi exposes whole-archive recall; Hermes search filtering does not constrain every get or SQL request. HTTP Host checks cover MCP and arbitrary SQL, not every JSON endpoint or caller authentication. SQL rejects corpus mutations but can write exported result files: RT-RTE-openclaw-tools, RT-CLM-transport-gates, RT-RTE-corpus-sql.
Restoration is a different consumer path
Restoration selects supported lineage, serializes canonical history and writes native client files; Pi's local interactive bridge then requests a host session switch. Existing destinations are refused through exclusive creation; unsupported deeper lineage is refused rather than partially restored. Failed-batch cleanup is best effort, not crash-atomic rollback. URL mode does not provide that local Pi bridge: RT-RTE-session-restore, RT-BAP-session-switch.
Retention of a checkpoint does not mean every restored consumer receives it. Pi native replay requires v3 origin, a Native request, complete raw records and captured placement. V4/SQLite origins reconstruct v3 with Foreign fidelity; reconstruction drops System carriers, including upstream summaries, while retaining them in the canonical archive. Pond archives those summaries rather than performing their original compaction. Host parsing, checkpoint interpretation and actual resumed work remain unobserved: MEM-OBJ-checkpoint-carriers, MEM-OBJ-raw-record, memory restore annotation.
Updates, checks and what they establish
The strongest supported trace-fed contribution is automatic production of durable conversational search projections and message vectors for later retrieval. These are genuine artifact-update chains, not merely stored raw logs. The verified comparison treats them as wired trace learning, independently of novel knowledge, parameter training or demonstrated improved capacity. Retained inputs shape the update; their outputs influence later ranking and context supply. This contribution remains supported even though the broader inventory is partial: EPI-RTE-search-projection, MEM-RTE-message-embedding, profile verification.
Other retained state supports software consumers rather than unsolicited model-context delivery. Validated row-map/cursor progress feeds restarted capture planning; embedding eligibility supplies retained text to numerical updates. Model-ID and storage diagnoses can cause refusal or selected repair, including replacing stale-model vectors. These internal supply and recovery paths should not be confused with automatic transcript insertion into the host. OpenClaw deletion reconciliation only reports erase/preserve classifications; it does not purge canonical rows. Physical old-version cleanup is not semantic withdrawal: MEM-RTE-sync-watermarks, EPI-RTE-maintenance-check, EPI-RTE-maintenance-recovery, MEM-RTE-deletion-detection.
Checks license narrow conclusions: structural admission, metadata conditions, transfer-key coverage or retrieval rank against nominated references. They do not accept historical claims as true. Within the inspected capture/recall/maintenance routes, no evidence-consuming acceptance of a new ampliative historical claim or subsequent lifecycle integration was established. This is not a repository-wide absence assertion. Imported explanations remain imported, and fluent retrieval is not production of warranted new knowledge: EPI-RTE-ingest-check, EPI-RTE-retrieval-benchmark, epistemic bounded conclusion.
The theory-builder conditions remain separate. The records expose localized shipped guidance, but do not establish a theory-building pathway's condition 1, localized formulated theory; condition 2, decisions depending on that theory's content; condition 3, content-directed criticism with changed reliance or revision; or condition 4, retained criticism shaping another round. Delivery of instructions or criticism inside history is insufficient for consumption or iteration. The adapter-development playbook prescribes evidence, fixtures, mappings, tests and a PR, not an observed criticism/revision loop. Criticism of a consumed theory improving future capacity is independently unestablished. Reflection beyond the recorded machinery diagnoses and recovery is not established; those routes should not be inflated into reflective theory building. Autonomous theory building is also unestablished: automatic archive operations do not demonstrate computation performing every internal theory-building role. See RT-RTE-adapter-development, RT-RTE-session-capture, epistemic bounded conclusion.
For self-improvement at this boundary, diagnosis-directed changes to retrieval machinery are supported, but exercised improvement-directed self-change is not established. The records show machinery effects and potential later consumers, not subsequent operation causally depending on an objective-bearing update or successful improvement. This conclusion is independent of the positive trace-fed artifact updates. A host trace linking retained evidence to a revised answer, a repair trace linking diagnosis to changed subsequent retrieval, or a controlled downstream comparison would change the relevant assessments without automatically establishing the others: EPI-RTE-maintenance-recovery, MEM-RTE-message-embedding.
The source reports smaller first useful responses for five author-chosen queries on one corpus/machine, with tokens estimated as bytes/4 against median matching transcript sizes. Follow-up reads add cost. This supports a bounded reported first-read comparison, not a complete answer, general recall quality or causal task improvement. Current-pin raw runs, validated reference labels and independently varied retrieval components would be needed for stronger comparisons: MEM-CLM-context-cost, EPI-CLM-ranked-context, EPI-RTE-retrieval-benchmark.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Partial interactive recall-channel inventory, carried from report verification | SRC-1; RT-RTE-archive-recall, MEM-BAP-recalled-history, report-verification Limits | Requested tool workflow is traced; verification also identifies Pi's human-selected /pond editor-reference action |
No exhaustive interactive-channel comparison. Pasting an ID/metadata/optional snippet and get-session instruction establishes neither full-transcript push, submission nor model adoption | Accepted route records tracing selection, reference construction, editor insertion and subsequent submission/consumption |
| Codec and opaque-payload coverage is incomplete | SRC-1; MEM-OBJ-part, MEM-OBJ-raw-record, MEM-OBJ-checkpoint-carriers, EPI-RTE-field-extraction | Shared pipeline and selected Pi/OpenClaw paths; other codecs not individually audited | No exhaustive fidelity, checkpoint, payload-form or original-summary-control conclusion; cap and reconstructed-carrier omission remain concrete exceptions | Codec-specific mappings, writer evidence and round-trip artifacts; upstream summary-generation and rationale traces |
| Host activation and dependency operation are unobserved | SRC-1; RT-RTE-archive-recall, RT-RTE-session-restore, RT-BAP-session-switch, RT-CLM-ingest-recovery | Frozen shipped integration/store code, not external hosts or storage internals; no execution performed | No reliable model compliance, successful continuation, observed convergent retry or deployed reliability/security conclusion | Scoped host and storage execution traces, failure/retry evidence and deployment contracts |
| Embedding identity and diagnosis are limited | SRC-1; RT-CMP-candle-embedder, MEM-OBJ-message-vector, EPI-RTE-maintenance-check | Loaded inference and model-ID metadata, including sampled diagnosis | No exact external weight provenance, same-ID drift detection or proof of corpus-wide model homogeneity | Immutable model revisions/digests and complete relevant corpus metadata checks |
| Knowledge, capacity and self-improvement conclusions lack their required links | SRC-1; RT-RTE-adapter-development, EPI-RTE-maintenance-recovery, MEM-RTE-message-embedding, epistemic bounded conclusion | Inspected archival, recall, maintenance and development-doctrine paths; broader review/research routes incompletely inspected | No accepted novel historical claim, complete system-negative, theory-builder loop, improved future capacity or exercised self-improvement attribution | Claim-specific acceptance evidence; formulated criticism and next-round consumption; objective-bearing update and subsequent causally dependent operation, with outcome comparisons for benefit |
| Evaluation comparison is narrow | SRC-1; MEM-CLM-context-cost, EPI-CLM-ranked-context, EPI-RTE-retrieval-benchmark | Five-query attributed context-cost report and inspected reference-matching evaluator, not current-pin raw runs/full label audit | No universal savings, final-answer completeness, answer correctness or component-level causal benefit | Current-pin runs with explicit corpus/query conditions, validated labels, full follow-up costs and controlled outcome comparisons |
Profile verification declares no additional limits. Source-read log: none; this synthesis uses the supplied members and adds no source evidence.