AgenticMarketFreshContext is a temporal correction layer for retrieval systems.
It exists to fix one specific bug: AI agents quoting stale information with full confidence.
The problem
Most retrieval pipelines rank by semantic similarity. Cosine similarity knows whether a document matches a query, but it does not know whether the document is still current.
So a 2022 blog post and a 2026 paper can score nearly identically. The model receives stale context and faithfully summarizes outdated information for a current question.
That is not always hallucination.
It is often correct summarization of corrupted retrieval.
The fix
FreshContext applies a decay-adjusted relevance score before context reaches the model:
R_t = R_0 · e^(−λt)
Where:
- R_0 is the base semantic relevance score
- λ is a source-specific decay constant
- t is time since publication
- R_t is relevance adjusted for freshness at query time
No model swap. No re-embedding required. FreshContext can sit as a post-retrieval reranking layer on top of an existing retrieval pipeline.
Live demo
No API key required:
https://freshcontext-mcp.gimmanuel73.workers.dev/demo
The demo uses the same model, same query, and same retrieval set. Only temporal correction changes.
A 2022 LangChain blog falls from rank 1 to rank 5. A 2026 X post rises from rank 4 to rank 1. The model receives different context, so the final answer changes.
When this helps
FreshContext is useful when the answer changes over time:
- “What is the current best practice for X?”
- “Is library Y still actively maintained?”
- “What did company Z just announce?”
- “Is this job still open?”
- “What is the latest research on X?”
When this does not help
FreshContext is not useful for timeless knowledge.
It does not improve questions like:
- “How does compound interest work?”
- “What is X?”
- “Explain how Y works.”
If the truth does not decay, freshness ranking is not the tool you need.
The standard
FreshContext wraps retrieved content in a structured envelope:
[FRESHCONTEXT] Source: https://github.com/owner/repo Published: 2024-11-03 Retrieved: 2026-03-05T09:19:00Z Confidence: high
... content ... [/FRESHCONTEXT]
This gives agents the three things raw retrieval usually hides:
- where the information came from
- when it was published
- when it was retrieved
The FreshContext Specification is published as an open standard under the MIT license.
What is deployed
FreshContext currently ships 21 reference tools across:
- GitHub repositories
- Hacker News
- Google Scholar
- arXiv
- Y Combinator companies
- Product Hunt launches
- npm and PyPI package trends
- no-key finance quotes via Stooq
- remote jobs
- SEC filings
- government contracts
- GDELT global news
- composite landscape tools
The adapters are demonstrations of the layer.
The layer is the product.
Production notes
The reference deployment runs on Cloudflare Workers and includes:
- MCP JSON-RPC transport
- live demo endpoint
- decay-scored intelligence feed
- D1-backed watched-query storage
- KV-backed rate limiting
- provenance signatures
- read-only safe-mode tools
Links
Live demo: https://freshcontext-mcp.gimmanuel73.workers.dev/demo
Website: https://freshcontext-site.pages.dev/
GitHub: https://github.com/PrinceGabriel-lgtm/freshcontext-mcp
npm: freshcontext-mcp
License: MIT
Built by a solo developer who got tired of AI agents confidently quoting two-year-old job postings.