Hippocampus boundary

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

Boundary and evidence

This run analyses the shipped Hippocampus framework for turning personal sources and conversation conclusions into a persistent, cross-linked Markdown vault and reading that knowledge in later agent sessions. The target is a memory/knowledge/context-engineering system, not an enclosing model runtime. Its model-dependent operations are specified by project instructions and skills for external Claude Code and Codex hosts. Its executable components provide lifecycle bookkeeping, deterministic structural checks and framework synchronization (SRC-1: CLAUDE.md, .claude/skills/ingest/SKILL.md, .claude/skills/save/SKILL.md, scripts/on-stop.sh, scripts/vault_lint.py, scripts/sync_framework.sh).

The user drops messy notes into inbox/; the agent turns them into a structured, cross-linked knowledge base under wiki/. --- CLAUDE.md

The boundary is whole-system for this repository's shipped framework. It includes both host adapters, canonical ingest/save/lint prompts, schema and page skeletons, original-source archiving, catalog/journal/cache upkeep, question-answer read order, session-start reload, stop-time persistence and the framework update/export path. It does not narrow the target to the Python linter or omit the prompts that produce and consume retained state. The source tree and shipped callers were traced from AGENTS.md to CLAUDE.md, from skill adapters to canonical skills, from host hook configuration to scripts/on-stop.sh, and from the lint skill to scripts/vault_lint.py (SRC-1: those paths and .agents/skills/ingest/SKILL.md, .agents/skills/save/SKILL.md, .agents/skills/lint/SKILL.md, .claude/settings.json, .codex/hooks.json, .claude/skills/lint/SKILL.md).

The later-consumption path is explicitly shipped, even though its execution belongs to the external host:

1. Read wiki/hot.md — recent context may already answer it.
2. Read wiki/index.md — pick the 3–5 relevant pages.
3. Read those pages; follow wikilinks at most one hop further. --- CLAUDE.md

CLAUDE.md also supplies a search fallback and instructions for another agent project to consult the vault. Both host configurations print the hot cache at session start and invoke the shared stop hook. The hook checks file modification times, can return a refresh request, and stages content roots before a local commit. These are implementation findings, not evidence that either host executed the configuration successfully (SRC-1: CLAUDE.md, .claude/settings.json, .codex/hooks.json, scripts/on-stop.sh).

if [ "$stop_active" -eq 0 ] && [ -f wiki/hot.md ]; then

--- scripts/on-stop.sh

External dependencies are the model and agent host, their instruction/skill loading, permissions, native file/web tools and hook protocol; Git; Bash and shell utilities; and Python for the linter. Obsidian is the optional browsing consumer. Their implementations are not part of this repository. Consequently the run can characterize the shipped contracts and hook logic but cannot establish host compliance, actual model extraction or retrieval quality, prompt-injection resistance, browser behavior, or compatibility across host versions. Framework synchronization code is in scope, but the live upstream revisions it would fetch are not evidence for this run (SRC-1: README.md, CLAUDE.md, .claude/settings.json, .codex/hooks.json, scripts/sync_framework.sh).

Tracked-root coverage

Every top-level tracked file or directory is classified below. “Included” denotes inspected source text unless explicitly identified as an empty scaffold. There is no material shipped loop path left uninspected. Every path in this table belongs to SRC-1.

The frozen evidence is repository commit 99531951ae60a9814ac7d27a32dd00f0c81a5008; the applicability cutoff is 2026-11-08. Overall evidence is code-grounded because hook, linter and synchronization implementations and their shipped drivers were inspected. Semantic ingest, save, retrieval and editorial maintenance remain instruction-grounded within that overall tier. No target execution was performed. The tracked tree supplies no test/evaluation suite or observed-run artifacts; the linter is a structural evaluator, with the lint skill supplying an editorial procedure, not a retrieval benchmark (SRC-1: scripts/vault_lint.py, .claude/skills/lint/SKILL.md, wiki/hot.md, wiki/index.md, wiki/log.md). These limits prevent claims of measured reliability, recall, latency, scalability or successful end-to-end operation; they do not prevent source-level analysis of the complete shipped framework.

Source register