Basic Memory boundary

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

Boundary and evidence

Basic Memory qualifies as a memory/knowledge/context-engineering system serving model-driven callers. Its intended use is persistent knowledge shared between users and AI assistants, rather than ownership of an enclosing general-purpose agent runtime. This classification rests on its declared Markdown knowledge interface and inspected MCP and host-extension entry points, not the README's promotional claims of never forgetting (SRC-1, README.md, src/basic_memory/mcp/server.py, integrations/pi/extensions/index.ts).

Your knowledge lives as Markdown files that both you and your AI can read, write, and search. --- README.md

The frozen artifact is the canonical repository at full commit a72751c5d626380dedeed30f55f7a80ba3dc710d. The acquired checkout's HEAD was checked against that pin without modifying it. The applicability cutoff is 2026-10-16, the supplied run date, not a claim about production deployment or the commit's authorship date. No additional repository, live documentation, external model service, or execution trace is registered.

Functional boundary and entry-point trace

The functional target includes the Python knowledge engine, its CLI/API/MCP surfaces, retained-content handling, shipped model-facing instructions and skills, first-party host integrations and plugins, and repository evaluators. It is not narrowed to the MCP library API. Host-native callers that persist and reintroduce retained content belong to this artifact even though they run inside another product (SRC-1, pyproject.toml, README.md, integrations/pi/package.json, integrations/pi/extensions/index.ts).

The inspected entry-point trace is:

1. basic-memory and bm resolve to the same CLI application. Its main module registers the MCP, hook, tool, import, project and maintenance command modules (SRC-1, pyproject.toml, src/basic_memory/cli/main.py).
2. The mcp command imports tools, prompts and resources and invokes the FastMCP server. The server lifespan initializes the application, starts the watch coordinator and drains pending materializations and background tasks at shutdown. This establishes maintenance and persistence responsibilities inside the target; it does not yet establish every storage invariant (SRC-1, src/basic_memory/cli/commands/mcp.py, src/basic_memory/mcp/server.py).
3. The inspected write_note tool constructs an entity and delegates to the project-scoped KnowledgeClient.write_note. build_context delegates to MemoryClient.build_context and formats or serializes the returned graph for callers. The trace reaches these API-client boundaries, not a complete inspection of the receiving API, database transactions or materialization workers (SRC-1, src/basic_memory/mcp/tools/write_note.py, src/basic_memory/mcp/tools/build_context.py).
4. The Pi package declares its extension entry point. That extension captures session turns through CLI tool write-note, searches earlier pi_session notes for recall, and can use hook session-start and hook pre-compact. Its before_agent_start handler returns recalled content as a message. This is an inspected shipped consumer of retained content, not an imagined external caller (SRC-1, integrations/pi/package.json, integrations/pi/extensions/index.ts).

[project.scripts] basic-memory = "basic_memory.cli.main:app" bm = "basic_memory.cli.main:app" --- pyproject.toml

const content = cfg.useHookFlow ? await runHook(cfg, ctx, "session-start", "startup", ctx.signal) : hasProjectMapping() ? await buildRecall(cfg, event.prompt, ctx.signal) : SETUP_GUIDANCE; if (!content) return; return { message: { customType: ENTRY_TYPE, content, display: true } }; --- integrations/pi/extensions/index.ts

Model-facing maintenance instructions are also part of the system, not merely development documentation. The inspected lifecycle skill instructs agents to archive notes and update their frontmatter. This supports an account of prescribed maintenance, not evidence that users or agents follow it (SRC-1, skills/memory-lifecycle/SKILL.md).

Manage how entities move through status stages in Basic Memory. The core principle: archive, never delete. Completed work is valuable context — move it out of the active view, but keep it in the knowledge graph. --- skills/memory-lifecycle/SKILL.md

Coverage status and prevented conclusions

The boundary kind is complete artifact, partial loop. The whole pinned repository is available as evidence, but this initial inspection does not demonstrate whole-system loop coverage. The following are inspection limits, not access failures or a prohibition on later analysts inspecting additional files at the same pin.

All rows refer to SRC-1. An uninspected available path is not an access gap. Newly inspected paths at the registered commit may extend evidence coverage without adding a source identity.

External dependencies and exclusions

Python dependencies, including FastMCP, FastAPI, SQLAlchemy, SQLite/vector tooling and optional backend/provider packages, are declared in pyproject.toml (SRC-1). Their separate implementations are outside this source set; this prevents assigning their internal guarantees to Basic Memory from declarations alone. Pi itself and its MCP adapter are external dependencies, while the Basic Memory Pi extension is included. The extension's imports and runtime registration establish that integration seam, not the host's scheduling or approval semantics (SRC-1, integrations/pi/extensions/index.ts).

The README describes a hosted cloud offering and external client/editor integrations. Live cloud infrastructure, authentication providers, other clients' implementation code, user knowledge bases, and actual production sessions are not captured. Consequently this run cannot verify hosted availability, team isolation, real-world recall quality, client compliance with instructions, or user outcomes (SRC-1, README.md). These exclusions do not remove cloud-routing code or first-party integration code present in the pinned artifact from the functional target.

The overall tier is code-grounded: implementation and declared design are both available, but no target code, tests or examples were executed. No observed-run or causal-experiment evidence is supplied.

Source register