Hindsight boundary

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

Boundary and evidence

This run examines Hindsight as a memory/knowledge/context-engineering system. Its intended use is an evidence-grounded account of memory production, maintenance, admission and later consumption for design-space comparison. It does not evaluate Commonplace adoption or establish benchmark superiority.

The target is the repository artifact, not only its Python library API. Functional inclusions are the memory service, asynchronous maintenance, retained-state retrieval and reflection, shipped clients and embedded launchers, integration hooks, agent-facing instructions, administration and evaluators. A shipped caller remains inside the target when it selects content to retain or injects retrieved content into an agent's context. The external host's own model orchestration and task execution do not thereby become Hindsight responsibilities.

Hindsight's own declared purpose supports the memory-system classification (SRC-1, README.md):

Hindsight™ is an agent memory system built to create smarter agents that learn over time. Most agent memory systems focus on recalling conversation history. Hindsight is focused on making agents that learn, not just remember. --- README.md

This classification is also code-grounded, rather than solely a reading of its branding. The packaged service exposes HTTP and optional MCP over a supplied memory engine (SRC-1, hindsight-api-slim/hindsight_api/api/__init__.py). The reflection implementation issues model tool calls itself (SRC-1, hindsight-api-slim/hindsight_api/engine/reflect/agent.py):

result = await llm_config.call_with_tools(**ct_kwargs) llm_duration = int((time.time() - llm_start) * 1000) queued_ms = int(queue_wait.seconds * 1000) consecutive_errors = 0 --- hindsight-api-slim/hindsight_api/engine/reflect/agent.py

The initial trace follows the packaged hindsight-api and hindsight-worker entry points into the memory engine and API construction. The worker imports the engine, initializes it with a worker task backend and constructs a poller; the service has separate HTTP/MCP construction paths. These are shipped runtime responsibilities, not an excluded enclosing application (SRC-1, hindsight-api/pyproject.toml, hindsight-api-slim/hindsight_api/main.py, hindsight-api-slim/hindsight_api/worker/main.py, hindsight-api-slim/hindsight_api/api/__init__.py). Pipeline inspection initially covers retain orchestration, the consolidation module's declared maintenance contract and reflection's executable model-call boundary; it is not an exhaustive trace of storage, admission or recovery.

A shipped Claude Code caller provides a concrete later-consumption path. Its hook manifest schedules recall on UserPromptSubmit and asynchronous retention on Stop. The retain script submits transcript content, a document ID, metadata and tags through its client. The recall script calls recall, formats returned results and emits context for the host (SRC-1, hindsight-integrations/claude-code/hooks/hooks.json, hindsight-integrations/claude-code/scripts/retain.py, hindsight-integrations/claude-code/scripts/recall.py):

output = { "hookSpecificOutput": { "hookEventName": "UserPromptSubmit", "additionalContext": context_message, } } json.dump(output, sys.stdout) --- hindsight-integrations/claude-code/scripts/recall.py

This establishes a shipped injection interface, not evidence that a real host used the returned context or improved its decisions. The host runtime is an external dependency; the scripts and hook contract are included Hindsight evidence.

Extent of inspection

The boundary kind is complete artifact, partial loop. The complete repository at the frozen commit is available as evidence. Initial inspection establishes the service and one concrete caller, but does not trace every material shipped consumer, evaluator, administrative mutation or deployment mode. No missing file or source-access failure is asserted. Available but uninspected material is not an access gap and is not excluded from the registered Git source. Later jobs may inspect it at this same commit; the present boundary does not warrant a whole-system coverage claim.

This is an initial-inspection map, not the exhaustive top-level coverage table required for a whole-system boundary. Remaining tracked root metadata and assets are available at the same pin but have not been treated as evidence of memory behavior.

Frozen source, exclusions and dependencies

The frozen evidence is SRC-1 at full commit 610c135b6f2d0d8b972da94146ae635943f36271; the applicability cutoff is 2026-10-14, the supplied run date. The acquired checkout's HEAD was verified to equal that pin. The source object is retained unchanged. The cutoff records this run's evidence boundary, not an assertion that the repository describes every deployed instance on that date. No live documents, additional repositories, provider calls, tests or examples were executed.

Deliberate exclusions are external agents and framework implementations, hosted Cloud internals, live benchmark dashboards, linked papers and moving online documentation. Excluding external hosts prevents conclusions about their own scheduling, reasoning and actual response to injected context. Excluding hosted internals prevents conclusions about managed-service tenancy, backups and reliability. Excluding unregistered benchmarks and papers prevents independent verification of comparative accuracy or research claims. Frozen in-repository declarations remain attributable evidence, not independent confirmation (SRC-1, README.md).

External dependencies include model providers and database deployments, plus host hook/framework protocols where integrations use them. The README declares provider choices and PostgreSQL/Oracle deployment options; the inspected reflection code demonstrates the model-call interface. Provider behavior, database behavior and host implementation are not supplied execution evidence (SRC-1, README.md, hindsight-api-slim/hindsight_api/engine/reflect/agent.py, hindsight-integrations/claude-code/hooks/hooks.json). These limits prevent guarantees of real-world model quality, database reliability or host interoperability. They do not prevent source-level analysis of Hindsight's side of those interfaces.

The overall evidence tier is code-grounded. Module comments, agent instructions and README descriptions supply doctrine/design. Evaluation documentation supplies attributed reported operation only. No inspected execution trace or controlled intervention establishes observed operation or causal effects.

Source register