Recall (parcha-skills) boundary

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

Boundary and evidence

Intended use. recall is an agent skill for Claude Code, Codex and pi. It instructs the agent to retrieve prior coding-agent sessions from a session-history index, read only the best-supported part of a session, and then answer, continue, repeat or turn the found work into a skill. The skill drives an engine that a model-run agent calls; its operation consists of the model following the instructions. It therefore falls in scope as a narrower mechanism of agent memory retrieval and context engineering, not as a whole memory system (SRC-1, recall/skills/recall/SKILL.md).

Target and scope selection. The caller restricts the target to the recall/ subtree; the other skills in the repository are context only. The package declares itself instruction-only:

Agent instructions for using a separately installed Recall system from Claude Code, Codex, or pi. This package contains the skill and plugin manifests. The implementation, installers, and deployment configuration are maintained separately and are not included. --- recall/README.md

The Recall engine is distributed separately. This skill contains no executable implementation. --- recall/skills/recall/SKILL.md

The repository's own test enforces that the package holds only .md, .json, .yaml and LICENSE files (SRC-1, tests/test_portability.py, test test_recall_and_tether_packages_contain_only_instructions). The tree listing at the reviewed commit confirms eight tracked files under recall/.

Functional inclusions.

• The skill procedure in recall/skills/recall/SKILL.md: interface selection (Recall MCP tools versus the local RECALL_SCRIPT CLI), the cold-start raw transcript scan with rg, outcome selection (find/verify, continue, repeat, skill-ify), search filters, evidence assessment through the WHY line, bounded reading through show, related, session-export, session-relations, index health, and the documented gotchas.
• The remote "central Brain" usage procedure in recall/skills/recall/references/central-brain.md, including deliberate capture through recall_capture and deletion through recall_forget.
• The query guidance in recall/skills/recall/references/query-cookbook.md.
• Packaging and discovery metadata: recall/package.json, recall/.claude-plugin/plugin.json, recall/.codex-plugin/plugin.json, recall/README.md, recall/LICENSE, and the repository-level registrations that expose the skill (.claude-plugin/marketplace.json, .agents/plugins/marketplace.json, root package.json).

Functional exclusions and external dependencies.

• The Recall engine (recall.py), its SQLite index, ingest, ranking, entity index, redaction, retention watchdog and session-export/session-relations producers. They are not in this repository at any path. The skill names the engine as a separately and privately distributed dependency. This prevents any implementation-level conclusion about how session state is produced, indexed, ranked, redacted or exported. Every statement about engine behavior (exact identifier matching, the evidence gate, thinking blocks never being stored, secret-shaped line redaction, incremental indexing) is available only as the skill's declared contract (doctrine/design), not as inspected behavior.
• The central Recall Brain service and its MCP tools (recall_search, recall_show, recall_session_context, recall_exec, recall_capture, recall_forget). These are private and absent. This prevents conclusions about cross-device storage, write admission, visibility enforcement, import connectors and deletion semantics beyond what the skill instructs the agent to do.
• The host harnesses (Claude Code, Codex, pi) that load the skill and produce the transcripts under ~/.claude/projects and ~/.codex/sessions. Their transcript formats and retention (cleanupPeriodDays) are external inputs; the target does not own them, and their responsibilities are not attributed to it.
• The harness skill creator used by the skill-ify outcome. It is an optional external dependency; the skill gives only a fallback instruction to write a standard SKILL.md package.

Consumers inside the repository. Other shipped skills consume Recall output. recap calls the engine's session-export and session-relations commands and checks their schema versions (SRC-1, recap/skills/recap/scripts/recap.py):

if page.get("schema_version") != "recall.session-export.v1": --- recap/skills/recap/scripts/recap.py

precap instructs its agent to search prior sessions with Recall when installed (SRC-1, precap/skills/precap/SKILL.md). These consumers call the separately distributed engine, not code in recall/; the recall/ skill documents the export interface they rely on. They are excluded from the functional boundary by the caller's scope, and because their own consumption logic belongs to other skills. They may be inspected as context for how retrieved session evidence is later consumed. This exclusion prevents attributing Recap's or Precap's evidence handling to Recall; it does not hide any Recall responsibility, since the export producer is absent either way.

Boundary kind. The skill package is a complete artifact: every file it ships is in the inspected tree. The memory loop it participates in is partial: the paths that produce, maintain and admit session state (indexing, redaction, central capture) are outside the repository, and only the agent-side retrieval and consumption procedure is inspectable. Hence complete artifact, partial loop, not whole-system.

Frozen revision and cutoff. Git commit c82883b15e5e8f02483bee513b3c8dcac48e3c9e of https://github.com/Parcha-ai/parcha-skills, committed 2026-10-22. The analysis cutoff is that commit date; claims apply to the skill text at that revision.

Evidence tier. doc-grounded. The target's shipped artifact is instructions. The skill text is the executable procedure the agent follows, so it supplies implementation evidence for the agent-side procedure. Everything about the engine and service it calls is declared contract only. No observed run or causal experiment is available for Recall within this source.

Coverage of top-level tracked entries.

Unavailable evidence: the Recall engine and the central Brain service. Deliberate exclusions: the other skills, by caller scope. Not yet inspected: the excluded skills' internals, which do not bear on Recall's mechanism.

Source register