Eidetic boundary

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

Boundary and evidence

This run characterizes Eidetic as a shipped memory/knowledge/context-engineering system for coding agents, not as the embedding library alone or as an enclosing coding-agent runtime. Its intended use is to retain personal knowledge across sessions and provide isolated, project-attached topic corpora. The boundary includes the repository's prompts, hook callers, durable writers, derived indexes, context assembly, maintenance, retrieval, exports and evaluation machinery. The whole-system classification refers to that shipped product; it does not claim that every optional path is enabled in an installed deployment. SRC-1: README.md, skill/SKILL.md, skill/eidetic-base/SKILL.md, install.sh.

The source declares the two consumption modes:

• Personal memory (PUSH) — your own decisions, rules, and project context. --- README.md

• Topic bases (PULL) — an external corpus (API docs, a methodology, a book) you turn into an isolated base and attach only to the projects that need it. --- README.md

The shipped loop is not just an API called by an unspecified external application. The installer registers SessionStart context injection, asynchronous Stop signal extraction, maintenance and PostToolUse lifecycle collection. The Stop hook reads a host-supplied transcript, calls an extraction model and passes filtered signals to compound.py; that module updates eligible existing cards or writes new signals. remember.py separately persists full answer cards. Later injection reindexes files and assembles retained content into a rules file; MCP and the recall skill expose on-demand search. These same-repository callers and writers are included. SRC-1: install.sh, hooks/session-signals.sh, bin/compound.py, bin/remember.py, hooks/smart-memory-inject.sh, bin/assemble_context.py, mcp_server.py, skill/SKILL.md.

Pass to compounding logic (now searches up-to-date index)

printf '%s\n' "$RESULT" | python3 "$COMPOUND" "$(pwd)" 2>/dev/null || exit 0 --- hooks/session-signals.sh

Assemble context via Python (crash-guarded — B6/H4: must not kill injection)

python3 "$MEMORY_SYSTEM/bin/assemble_context.py" "$DB" "$RULES_FILE" "$(pwd)" || true --- hooks/smart-memory-inject.sh

Included maintenance extends beyond declared drift checks: pending ingest dispatches to M1 contradiction and M2 synthesis modules, while the shipped M3 transcript driver routes consolidation, acquisition and an independently gated agent lane. Their presence is not evidence of deployment activation or successful model judgments. SRC-1: bin/index_impl.py, bin/m1_contradiction.py, bin/m2_synthesis.py, bin/m3_hook.py, hooks/semantic-maintenance.sh.

if os.environ.get("EIDETIC_M3_DRIVER", "").strip().lower() not in ("1", "on", "true", "yes"): print(json.dumps({"m3_driver": "skip", "reason": "dark"})) return 0 --- bin/m3_hook.py

Topic-base creation, curated addition, indexing and attachment are included, as are the engine module and the Core-owned JSONL worker. The latter accepts logical index identifiers and resolves storage inside Core. The SDK consumer and source-specific applications described in the ownership document are outside this repository. The register freezes only Eidetic, not those applications. SRC-1: bin/base.py, mcp_server.py, bin/engine.py, bin/eidetic_engine_worker.py, schemas/sdk/engine/v1/contract-manifest.json, docs/core-sdk-ownership.md.

Repository coverage

Every top-level tracked entry at the reviewed commit is classified below. All paths are SRC-1 commit-relative anchors. “Included” means available to this analysis, not exhaustively inspected or executed. Initial inspection traced the production, persistence, reload, admission and consumption seams named below. Remaining internal detail is available for later jobs; it is not an access gap.

External dependencies and limits

Claude Code owns lifecycle event delivery, transcript production and loading the generated rules; other MCP agents own when to call tools and how to use results. Eidetic's shipped adapters are included, but those hosts are not frozen. This prevents conclusions about actual host compliance, end-to-end agent behavior or cross-host parity. SRC-1: install.sh, hooks/smart-memory-inject.sh, hooks/session-signals.sh, skill/SKILL.md, README.md.

Extraction and judgment depend on model routes and external CLIs/services; semantic retrieval depends on embedding/reranking models and optional packages, with additional translation and MLX paths. Python, Bash, SQLite, file locking and local host configuration are deployment dependencies. Only the dependency-facing code and declarations in SRC-1 are frozen. No provider service, model weights, installed runtime, user memory corpus, transcript stream, private recall battery, external SDK, importer or YouGile consumer was acquired. This prevents provider-quality, deployment-activation, private-corpus performance and external-integration correctness conclusions. SRC-1: README.md, hooks/session-signals.sh, bin/m3_hook.py, bin/recall_lab.py, bin/recall_smoke.py, docs/core-sdk-ownership.md.

The fixed evidence is commit 67506b62fdd0944c243537c3cb10938f7343b6b2, with an applicability cutoff of 2026-10-14. HEAD matched the acquired revision and the checkout reported no changes during boundary inspection. The overall tier is code-grounded: implementation and source-owned design are inspectable, but this job executed neither the target nor its tests and registered no observed-run or causal-experiment evidence. Available but unread files above are coverage work, not unavailable evidence.

Source register