Cognitive infrastructure for AI agents

Understand context.
Learn from experience.

DeepDream researches how AI agents interpret business context, update long-term memory, and apply past lessons to the next task.

FIG. 01 / Memory field
Meaning and time determine which experiences matter now.
DEEPDREAM

Technical architecture

From context to reusable experience.

Understand the concepts and constraints first. Then decide what to retain, when to recall it, and how to verify that it helps.

FEATURES / DREAM

Make the last task useful for the next.

Dream is DeepDream’s design for background memory maintenance. Between sessions, it reviews task logs, tool results, and feedback to update the memory an agent can read without changing model weights.

Design illustration · weekly active users (WAU) Four layers of one business memory
Startup context
Confirm the team and metric version before analyzing.
Knowledge representation
A metric page links both definitions, the change record, and query templates.
Memory notes
The growth team changed its criterion from login to a core action.
Raw traces
Keep the approval record and query results for later verification.
  • CORRECT

    Check memory against the evidence

    When a note conflicts with logs or verified results, distinguish an original mistake from a later rule change. Correct the mistake, retain dated versions for rule changes, and flag uncertain notes for review.

  • ORGANIZE

    Give scattered notes a structure

    Merge duplicates by entity or topic, then connect related notes through pages, indexes, and links. Keep source records searchable so an agent can check the original, not just the summary.

  • PRIORITIZE

    Preload only what matters

    Keep a short startup context with verified, frequently useful constraints and pointers. Load details on demand and move stale items out of startup context.

  • FOCUS

    Keep what serves the task

    For metric analysis, prioritize definitions, effective dates, and validation steps. Use task feedback to decide which rules are worth keeping rather than saving every action as lasting memory.

  • FILL GAPS

    Find what one session missed

    Compare sessions to uncover recurring mistakes and missing conditions. Add notes backed by evidence, repair broken links, and leave unsupported gaps unresolved.

Use case · DATA AGENT

When “active user” changes meaning

A growth team revises its definition of active users. When the next anomaly appears, the agent needs to know when and why the definition changed, and which analysis steps still apply.

Illustrative scenario · not production data

Memory decision trace

Metric: weekly active users (WAU)

Resolve the metric definition before running the analysis

Different teams can use the same metric name for different definitions. Identify the team and reporting period. If the definition is still ambiguous, ask rather than assuming the last one applies.

Team and time range

Confirmed change: login → core action

Record both versions and when each applies

Record the revised definition, its source of approval, and its effective date. Keep the old version for historical queries without treating it as current.

Evidence and effective dates

Recalled: a verified channel breakdown

Apply only the lessons that still hold

Check task fit, access, timing, and provenance. Keep the channel breakdown method, but reject the SQL template that still filters by the old definition.

Relevance and scope of use

Check: recompute the trend with one definition

Separate a definition change from a real business change

Recompute comparable periods with a consistent definition and record whether the evidence supports the conclusion. Save useful steps with their conditions so they can be tested again.

Outcome and feedback

DEEPDREAM RESEARCH

Remembering is not enough.

Explore how semantic modeling, the memory lifecycle, and feedback help agents use past experience on real tasks.

View the architecture