# Knowledge & memory

Help an agent continue its work, use past experience, and support decisions with evidence.

Author: Murali Sid (https://linkedin.com/in/muralisid)
Source: https://www.agenticarchitectureskills.com/memory
Published: 2026-09-13
Updated: 2026-10-15
Licence: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/
Cite as: Murali Sid. Knowledge & memory. Agentic Architecture Skills. https://www.agenticarchitectureskills.com/memory

---


An agent should not start every task from nothing. It also should not treat everything it has ever seen as current, trusted, or available to everyone. **Good memory helps it carry work forward and show what supports its decisions.**

<TeachingIllustration src="/images/memory/memory-map-v1.webp" title="Three questions before choosing a memory product" alt="Working context holds what the model needs now. Thread memory preserves the current case. Long-term memory can retain episodes, facts, and approved methods for future tasks." caption="Purpose, retention scope, and representation are different design choices. They can be combined." />

## Start with a task that continues tomorrow [#start-with-a-task-that-continues-tomorrow]

An agent is collecting evidence for an inspection. Today it finds a report, identifies a missing photograph, and asks for a site check. Tomorrow it should know what remains open. Next month a different investigation may benefit from the completed case, but should not inherit an unverified guess as an established fact.

A long chat transcript alone does not make these distinctions. Neither does putting every document into a vector database. We need to decide what belongs to this task, what may help future tasks, and what must remain traceable to a source.

## Follow the memory story [#follow-the-memory-story]

The main path has three parts. Start with the first and follow the continuation at the end of each page.

<Cards>
  <Card title="1. Understand what to remember" href="/memory/memory-types">
    Working context, thread memory, and long-term episodes, facts, and procedures. See how they cooperate in one investigation.
  </Card>

  <Card title="2. Make it usable as evidence" href="/memory/information-to-memory">
    Capture sources, choose representations, combine observations, and handle changing facts and permissions.
  </Card>

  <Card title="3. Test the design" href="/memory/what-the-benchmarks-measure">
    Check retrieval, answers, and completed work separately. Then investigate cost and performance at scale.
  </Card>
</Cards>

## Go deeper when the question calls for it [#go-deeper-when-the-question-calls-for-it]

The sidebar separates the main story from optional visual tours and advanced design choices. You do not need to read every embedding article to understand memory. The [vector-space tour](/memory/vector-spaces) explains the different information forms together; its individual articles offer more detail.

The advanced pages compare storage designs, long context, large histories, and research findings. They are choices to investigate, not a list of components every enterprise needs.

<Cards>
  <Card title="Token efficiency: spend tokens on outcomes" href="/memory/token-efficiency">
    Why agents use many times the tokens of chat, what to meter for each agent, and which savings have been measured.
  </Card>
</Cards>

Memory supplies evidence within the [enterprise architecture](/architecture). The memory story follows [the data-to-memory pipeline](/architecture/data-to-memory) and belongs to the [build step on the Start page](/start#5-build). The [Intelligence ladder](/ladder) explains how an agent uses that evidence to do useful work at an acceptable cost.

**Next step:** start the story with [what an agent should remember](/memory/memory-types).