The map of context patterns

An agent’s answer is bounded by what the model reads before it answers. “Give the agent knowledge” sounds like one problem, but it is six different questions, and most failures come from answering one of them with a pattern built for another: chunk retrieval over a curated policy set, a vector store standing in for memory, or a bigger window standing in for a choice.

This page lays out the six questions, the patterns that answer each, and when each pattern wins. It is the map the Astromesh libraries are built against.

Question	Family	In Astromesh
Which fragment helps?	Retrieval	RAG
Where do I go next?	Navigation	OKF
How do things relate?	Graphs	Not yet
What have I learned?	Memory	Core runtime
What fits in the window?	Context engineering	Context
Can I trust it?	Provenance	OKF signals, honored by Context

Which fragment helps? Retrieval

The corpus is large and unstructured, and the answer is somewhere in it. Retrieval splits it into chunks ahead of time and finds the closest ones at question time.

Naive vector search wins when questions are worded like the documents.
Hybrid (vector plus BM25, fused) wins on exact tokens an embedding blurs: error codes, SKUs, version numbers, names.
Multi-query and HyDE win when users phrase things differently from the documents, at the cost of a model call before the search.
Reranking with a cross-encoder sharpens any of them, at the cost of one more model pass.
Contextual retrieval (prepending a sentence of document context to each chunk before embedding) and parent-document retrieval (search small chunks, return the larger section) fix the commonest failure: a chunk that is relevant but meaningless on its own.

Retrieval loses when the knowledge is small, curated and linked: chunking cuts through the structure the authors built, and the model reads fragments of a policy instead of the policy. RAG ships the first four as wirings of the same parts, with a trace per stage, and the patterns page explains how to choose.

The knowledge is curated: policies, runbooks, metric definitions, a data catalog. A person would not search it; they would open the index and follow links. Navigation gives the agent tools to do the same.

The index in the system prompt. The table of contents does not change between questions, so it belongs in the stable prefix, where it saves a tool call and caches well.
Agentic search with tools. Find, read, follow a link. The model reads whole documents and can cite the one it answered from.
Progressive disclosure. Load a one-line description of everything and the full body only when it is needed. Agent skills and llms.txt are the same idea.

Navigation loses on thousands of unstructured documents with nothing to follow, and every step is a model turn, which adds latency. OKF implements it over Open Knowledge Format bundles, where every concept carries who verified it and when it goes stale.

How do things relate? Graphs

Some questions are about connections rather than content: which services depend on this database, what changed between two contract versions, which customers share an account manager. Neither chunks nor documents answer them; the answer is a path.

Entity knowledge graphs extract entities and relations and answer by traversal.
GraphRAG clusters the graph into communities and summarizes each, which answers corpus-wide questions (“what are the main themes?”) that chunk retrieval cannot.
Temporal graphs add validity intervals to each fact, so “what was true in March” has an answer.

Graphs are expensive to build and keep correct, since extraction needs a model over the whole corpus. They pay off when the questions are relational. Astromesh has no graph library yet.

What have I learned? Memory

Memory is knowledge the agent produces rather than reads: what this user prefers, what was decided last week, what failed the last time.

Conversational memory keeps the recent turns, trimmed by a sliding window, a summary, or a token budget.
Semantic memory extracts durable facts (“the customer is on the Team plan”) and retrieves them by similarity.
Episodic memory logs what happened, so the agent can recall similar past situations.
Self-managed memory gives the agent tools to write, edit and forget its own memory, instead of storing everything automatically.
Consolidation periodically merges, deduplicates and retires facts, so memory does not become a second unstructured corpus.

The common failure is treating memory as retrieval over raw transcripts: it grows without bound and retrieves contradictions. The Astromesh core implements the first three; to Context, memory is one more source competing for the window.

What fits in the window? Context engineering

Every family above produces candidates. The model has one window, and long windows are not free: cost and latency grow with every token, and models attend unevenly to long contexts (the “lost in the middle” effect). Something has to choose.

A budget across sources. Fixed slices per source waste tokens: when retrieval finds nothing useful, its slice fills with noise while memory gets cut. A shared budget on comparable scores spends tokens where they help.
Deduplication. Different sources return the same content more often than expected; a retrieved chunk and a navigated concept are frequently the same policy.
Cache-stable ordering. Providers cache the longest shared prompt prefix. Putting stable content first and keeping it byte-identical turns repeat calls into cache hits.
Compaction. Long-running agents summarize old turns and keep recent ones verbatim, without splitting a tool call from its result.
Offloading. A large tool result goes to storage and leaves a handle; the agent reads the part it needs.
Isolation. A sub-agent explores in its own window and returns only its conclusion.

This is the family most stacks leave implicit, spread across templates and trim functions. Context is the Astromesh library for it, and its eval compares policies on accuracy, tokens and prefix reuse instead of asserting that one is better.

Can I trust it? Provenance

Content in the window is a claim. The model needs to know where it came from, whether anyone checked it, and whether it still holds.

Citations. Every item keeps an identifier the answer can point to.
Freshness. Content past its review date is marked stale rather than silently trusted.
Verification. Human-reviewed content is distinguished from generated or imported content.
Conflict handling. When two sources disagree, the conflict is surfaced rather than resolved by whichever came last.

The principle Astromesh follows is from the OKF spec: signals are annotated, never used to hide content. OKF returns trust, staleness and deprecation with every concept, and Context carries them into its scoring and its trace.

Combining them

The families are layers, not alternatives. A support agent might navigate a curated policy bundle, retrieve from a large ticket archive, remember the customer’s plan, and let a context policy decide which of those the model reads on each turn. The first three families decide what could be relevant; context engineering decides what is read; provenance travels with all of it.