Skip to main content

Research Reports

schema.org/Report — structured research notes with citations, knowledge graph annotations, and provider metadata.

Research is a first-class workflow in the knowledge layer. When the agent investigates a topic — comparing tools, evaluating architectures, gathering competitive intelligence — the output is a Report entity with a consistent internal structure.

Example

---
type: schema.org/Report
id: e5f6a7b8-9012-34ab-cdef-567890123def
permalink: research/agent-memory/act-r-decay-models
title: "ACT-R Memory Decay in Agent Architectures"
about: applying ACT-R cognitive decay models to AI agent memory systems
mode: adaptive
status: completed
providers:
exa:
tool_calls: 3
model: exa-research-fast
tags:
- research
- memory
- act-r
- cognitive-science
---

# Instructions

## Context

We're designing a decay model for Attaché's memory system and need
to understand how ACT-R's base-level activation applies to agent
memory retrieval — particularly the power law of forgetting and
how frequency/recency interact.

## Questions

1. How does ACT-R's base-level activation formula work?
2. What optimizations exist for computing activation at scale?
3. How have other agent systems adapted ACT-R for memory decay?

# Report

## Base-Level Activation

ACT-R models declarative memory using base-level activation, where
each memory chunk's accessibility depends on how often and how
recently it's been accessed. The formula combines a logarithmic
sum of recency-weighted access events with a decay parameter,
producing a power law of forgetting that matches human recall data.

Anderson & Lebiere's work at Carnegie Mellon established that both
frequency and recency matter — a note accessed 50 times decays
much slower than one accessed once, even at the same age...

## Buffer Decay Extensions

Thomson, Bennati & Lebiere (2014) extended ACT-R's short-term
memory model with buffer decay, showing how spreading activation
provides implicit contextual information...

# Annotations

## Observations

- [finding] ACT-R base-level activation combines frequency and recency via power law #act-r #memory
- [finding] Optimized O(1) approximation avoids iterating over access history #performance
- [recommendation] Apply metabolic rate multipliers for different memory layers #architecture

## Relations

- informs [[projects/attache]]
- related_to [[research/agent-memory/ori-mnemos-review]]

## References

- [Anderson & Lebiere — The effect of memory decay on predictions from changing categories](http://act-r.psy.cmu.edu/?p=13636&post_type=publications)
- [Thomson et al. — Extending the Influence of Contextual Information in ACT-R using Buffer Decay](http://act-r.psy.cmu.edu/wordpress/wp-content/uploads/2015/09/CogSci-2014-Final-Extending-the-Influence-of-Contextual-Information-in-ACT-R-using-Buffer-Decay.pdf)
- [Ori-Mnemos — Agent memory with ACT-R and graph-aware extensions](https://github.com/aayoawoyemi/Ori-Mnemos)

Fields

FieldTypeDescription
titlestringReport title (required)
aboutstringSubject matter (Schema.org field)
modeenumdeep_research, adaptive, or external
statusenumdraft, pending, completed, or failed
providersRecordProvider metadata: tool calls, model, cost
tagsstring[]Classification tags

Document Structure

Research notes follow a consistent internal structure:

  1. Instructions — context, scope, and specific questions (written during draft phase)
  2. Report — the actual research output (written during execution)
  3. Annotations — knowledge graph metadata:
    • Observations — structured annotations extracted from findings
    • Relations — links to related entities via [[wiki-links]]
    • References — source citations with URLs

Research Modes

ModeHow it worksExample providers
deep_researchAutonomous report generation — provider executes independently and returns a complete reportExa Deep Research, Perplexity
adaptiveAgent-driven multi-tool exploration — the agent calls search, scrape, and documentation tools in sequenceFirecrawl, Exa search, GitHub CLI, Apollo
externalImport from external AI tools — user conducts research elsewhere and imports the resultsClaude Deep Research, ChatGPT share links

Draft-First Workflow

Research follows a draft-first pattern where the plan is written to the knowledge base before execution:

  1. Scope — Agent asks clarifying questions to refine the topic
  2. Draft — Agent writes a note with status: draft containing instructions and questions
  3. Review — User reviews, adds inline feedback
  4. Execute — User selects provider/model, research runs with finalized instructions
  5. Persist — Report, annotations, and references are written to the note

Provider Selection Guide

Query typeRecommended providerWhy
Semantic/conceptual questionsExa Deep SearchNeural understanding, query expansion
Exact keywords, error codesFirecrawlSearch operators, keyword matching
Known URL extractionFirecrawl scrapeBest content extraction
People and organizationsApolloStructured business data
Library/API documentationRefOptimized for docs
GitHub code and issuesGitHub CLIDirect access

Folder Structure

knowledge/research/
├── ai-agents/
│ ├── coding-agent-taxonomy.md
│ └── agent-memory-patterns.md
├── mcp-tools/
│ └── search-quality-comparison.md
└── infrastructure/
└── colima-vs-docker-desktop.md
Zod schema
export const ReportSchema = z.object({
type: z.literal("schema.org/Report"),
title: z.string(),
about: z.string().optional(),
mode: z.enum(["deep_research", "adaptive", "external"]).optional(),
status: z.enum(["draft", "pending", "completed", "failed"]).optional(),
providers: z.record(z.object({
tool_calls: z.number(),
model: z.string().optional(),
cost: z.number().optional(),
})).default({}),
tags: z.array(z.string()).default([]),
});