CONCEPT Cited by 1 source
Context file freshness¶
Definition¶
Context file freshness is the discipline of keeping AI-agent context files in sync with the underlying code they describe so that references (file paths, function names, invariants) remain accurate. The load-bearing observation Meta names explicitly:
"Context that goes stale causes more harm than no context. Build periodic validation and self-repair." (Source: sources/2026-04-06-meta-how-meta-used-ai-to-map-tribal-knowledge-in-large-scale-data-pipelines.)
Stale context isn't merely less useful — it's actively harmful.
An agent consuming a context file that refers to foo.py::parse()
after parse() was renamed to parse_v2() will confidently generate
code calling the non-existent function. The agent is worse off than
if the context file hadn't been loaded at all, because it "knows"
something that isn't true.
Why the asymmetry¶
| No context | Stale context |
|---|---|
| Agent explores + guesses | Agent confidently wrong |
| 15-25 tool calls wasted | Code produced that looks right but isn't |
| Visible slowness | Invisible incorrectness |
The failure mode with stale context is silent wrong output — the same failure mode config-as-code pipelines already suffer from. Layering stale context on top amplifies rather than mitigates.
Meta's quantitative stake¶
- Meta's invariant: zero hallucinated file paths across 59 context files (validated by final critic agents).
- Refresh cadence: "every few weeks" — driven by an automated loop.
- Files: 59, each ~1,000 tokens → total surface ~59,000 tokens to keep fresh against a 4,100-file / 4-repo / 3-language pipeline.
The tractability argument: compass-shaped files make freshness automatable because 25-35 lines per file are small enough for a critic agent to re-validate cheaply.
The self-refresh mechanism¶
patterns/self-maintaining-context-layer — Meta's canonical wiki instance:
- Validate file paths against the live repos (detect renames, moves, deletions).
- Detect coverage gaps (new modules added since last refresh).
- Re-run critic agents against updated content.
- Auto-fix stale references.
"The AI isn't a consumer of this infrastructure, it's the engine that runs it."
Why compass shape matters here¶
Compass-not-encyclopedia and freshness are interlocked:
- Encyclopedia-sized files (300+ lines) — freshness is intractable; one refactor invalidates many paragraphs, humans can't keep up, agents can't cheaply re-validate.
- Compass-sized files (25-35 lines) — each file is small enough that an automated critic agent can re-validate it in a single LLM call; mass refresh across 59 files fits inside "every few weeks".
Contrast with traditional documentation rot¶
Traditional documentation rot is a slow engineering-culture problem: wikis drift, nobody updates them, onboarding papers over the drift. The failure mode is humans reading outdated docs.
AI-agent context-file staleness is a fast, high-leverage problem:
- Amplification — one stale context file is consumed by every agent task that touches that module, possibly thousands of times between refreshes.
- Invisibility — agents don't protest stale docs; they act on them.
- Compounding — AI-assisted commits built on stale context push the code further from the context, accelerating drift.
The operational response has to match the pace: automated, frequent, gated.
Seen in¶
- Meta AI Pre-Compute Engine (2026-04-06) — canonical wiki instance. Self-refresh runs "every few weeks" with four automated steps; zero hallucinated file paths is an enforced invariant. (Source: sources/2026-04-06-meta-how-meta-used-ai-to-map-tribal-knowledge-in-large-scale-data-pipelines.)
Related¶
- concepts/tribal-knowledge — the content the fresh context carries
- concepts/compass-not-encyclopedia — the format that makes freshness tractable
- concepts/context-engineering — the parent discipline
- patterns/precomputed-agent-context-files — the pattern where freshness is first-order
- patterns/self-maintaining-context-layer — the automated refresh-loop pattern
- systems/meta-ai-precompute-engine — the canonical producing system