Skip to content

Context cost & lean delivery

A knowledge server justifies itself only if it stays lean — otherwise it becomes the noise it was meant to cure. Two findings shape how Lore delivers knowledge to an agent:

  • Context rot. Every frontier model degrades as input grows, well before the window fills — so dumping a corpus at an agent actively worsens output.
  • The MCP "context tax." A tool surface spends token real estate on descriptions and schemas before it answers anything; oversized surfaces have been cited around ~23k tokens.

Lore's answer is three deliberate properties: a measured, budgeted tool surface, selective on-demand retrieval by default, and a CLI delivery path that spends no standing tokens at all. None of them uses AI summarisation or compression — payloads stay small because they are scoped, not lossily shrunk (ADR-066).

1. The MCP surface is measured and budgeted

The standing cost of the MCP server — the five tool descriptions and their JSON schemas a client loads every session — is measured deterministically and offline (no model, no network) and held under a budget as a regression check. Today it measures ~915 tokens against a 1000 budget — roughly 25× under the ~23k figure the context-tax critique cited. A description or schema edit that inflates the surface fails CI rather than quietly taxing every session (see the rac-mcp-surface-budget requirement and rac/mcp/surface.py).

2. Retrieval is selective and on-demand by default

Both delivery surfaces return the relevant artifacts for a query, never the whole corpus:

  • search_artifacts / rac find return the matches for a query — a small, scoped result set, ranked and bounded by the response budget (ADR-033).
  • get_artifact / rac resolve return one artifact by id.
  • get_related / rac relationships return an artifact's immediate neighbours, not a transitive dump.

This is the antidote to context rot: an agent receives small, relevant payloads by construction, and pulls more only when it asks. Bulk, whole-corpus delivery is an explicit, opt-in action — rac export — never a retrieval default. No default path hands an agent the entire corpus; it must request it.

3. The CLI is a first-class, lowest-tax delivery path

The MCP server is not the only way to ground an agent, and it is not always the leanest. A CLI spends no standing token tax — it costs nothing until invoked — which is why the context-tax critique steers toward CLI utilities. Lore's CLI-first posture (ADR-005) makes this a supported choice, not an afterthought: find, resolve, and relationships deliver the same grounding the MCP tools do.

Ground an agent through the CLI by having it shell out and read JSON:

# What did we already decide about deleting users?
rac find "delete user" rac/ --json

# Read the specific decision the search surfaced.
rac resolve RAC-01JY4M8X2QZ7 rac/ --json

# Which recorded decisions govern the file I'm about to edit?
rac decisions-for src/users/repository.py rac/ --json

# What else would this change affect?
rac relationships rac/ --json

Each returns a small, scoped, JSON payload the agent can act on — the same selective-by-default retrieval as the MCP path, with zero standing surface cost.

When to use which

Both surfaces are first-class and supported; the choice is a context-cost trade-off, not a deprecation.

Prefer the CLI path when… Prefer the MCP server when…
The agent can shell out (CI jobs, scripted agents, terminal tools) The agent calls tools autonomously mid-conversation
You want zero standing token cost — nothing is paid until a command runs You want the model to decide when to retrieve, from the tool descriptions alone
Grounding is a discrete step in a pipeline Grounding is interactive and continuous

The MCP server stays fully documented and supported — see MCP Server. The CLI commands are in the CLI Reference. Use whichever spends the context you can afford; neither is deprecated in favour of the other.