Smriti MCP boundary

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

Boundary and evidence

This run analyses Smriti MCP as a memory/knowledge/context-engineering system serving model-driven agents. Its intended use is to preserve facts, decisions, preferences and reusable project context beyond a chat. It qualifies through the memory operations it serves over MCP, not through ownership of a model-call loop (SRC-1, src/smriti_mcp/server.py, README.md). The server's shipped instructions state:

"Smriti gives agents durable markdown memory. Store facts, decisions, " "preferences, project context, and reusable notes that should survive " "beyond the current chat." --- src/smriti_mcp/server.py

The boundary is whole-system for the shipped Smriti artifact, not for an enclosing agent application. Inspection traced the package command to the CLI, CLI server startup to create_server, tool wrappers to MemoryStore, and storage, reload, retrieval, trace consolidation and maintenance paths through their implementations. The Python library examples and test callers are included, not treated as external merely because they use the library API (SRC-1, pyproject.toml, src/smriti_mcp/cli.py, src/smriti_mcp/server.py, src/smriti_mcp/store.py, README.md, tests/test_smriti_store.py). The installed entry point is:

[project.scripts] smriti-mcp = "smriti_mcp.cli:main" --- pyproject.toml

Included functions are memory creation, metadata admission, append/update, archive/delete, filesystem reload, list/search, task recall, access reinforcement, raw trace recording and loading, consolidation suggestions and writes, supersession, health reports, frontmatter repair, wikilink maintenance, and markdown/YAML index generation and markdown index loading. These cover production, maintenance and subsequent consumption of retained state within the artifact. Agent-facing instructions and parameter descriptions in the server, host configuration examples, maintenance instructions and shipped tests are included (SRC-1, src/smriti_mcp/server.py, src/smriti_mcp/store.py, src/smriti_mcp/frontmatter.py, src/smriti_mcp/fix_frontmatter.py, QUICKSTART.md, FRONTMATTER_FIX.md, tests/test_smriti_mcp_integration.py). Recall's returned context is explicitly materialized for a caller:

return { "query": query, "count": len(memories), "memories": memories, "context": "\n".join(markdown_lines).rstrip() + "\n", "token_budget": token_budget, } --- src/smriti_mcp/store.py

The semantic decision to retain information, final consolidated prose, invocation timing and use of returned context in a model's next prompt belong to an external agent or operator. They are not attributed to Smriti. The shipped consolidation tool states that division:

"Create, append, or update a durable memory from agent-supplied consolidated content. " "Use after reviewing traces; Smriti records provenance but does not summarize content itself." --- src/smriti_mcp/server.py

External dependencies include Python, an accessible local filesystem, MCP/FastMCP for transport and tool serving, Pydantic for input models, and PyYAML for frontmatter/index serialization. MCP-compatible hosts such as Claude Desktop, Cline and Cursor are documented consumers; their implementations, model providers and prompts beyond the shipped examples are outside this repository boundary. Their exclusion prevents conclusions about autonomous capture policy, actual prompt injection of recall, cross-session task improvement or end-to-end agent reliability. Dependency internals are not frozen here, so source inspection supports Smriti's calls and declarations, not a full transport or dependency security assessment (SRC-1, pyproject.toml, requirements.txt, QUICKSTART.md, src/smriti_mcp/server.py, src/smriti_mcp/frontmatter.py).

The frozen source is the acquired repository at 8264bf9103477d9521e57b27c81dcc2e943a3230, with applicability cutoff 2026-10-14. The overall tier is code-grounded. Implementation and declared intent are distinguishable; README test-success claims and test source do not establish an observed run. No target code, tests, examples, providers or services were executed. There are no supplied execution traces, production memory corpora or causal experiments, so the boundary supports source-level mechanisms and test intent, not measured retrieval quality, performance, deployed correctness or causal benefit.

Shipped-path coverage

All rows concern SRC-1 at the frozen commit. Included executable files and tests were read in full. Documentation was inspected as declared intent, not as an instruction to execute. No material shipped memory-loop path remains uninspected. Exclusions below are deliberate, not access failures; no source access gap was encountered.

Source register