I'd like to propose an optional experience-memory layer for applications built with the Perplexity API.
The idea goes beyond storing conversation history. When an agent performs research, it often discovers useful information, failed approaches, source preferences, and conclusions that could be valuable to a future agent or research session.
Instead of making the next agent repeat the same research from scratch, validated experience could be transferred:
Research Agent A
↓
researches a topic
↓
sources + findings + failed approaches
↓
validated experience
↓
Memory
↓
Research Agent B
↓
related question
↓
retrieves relevant experience
↓
continues from previous work
For example, Agent A might discover that several sources are outdated, identify the most reliable sources, and verify a particular finding.
Instead of transferring the entire research conversation, the memory could preserve:
Research question
Relevant findings
Sources and provenance
Failed research approaches
Why a source/approach was rejected
Verification status
Timestamp
Relevant context
Confidence
I think failed research experience is particularly valuable.
Knowing "this source was outdated as of X" can prevent another agent from repeating the same mistake.
The receiving agent could first check whether the stored experience is still applicable before relying on it:
Stored experience
↓
Provenance check
↓
Freshness/context check
↓
Relevant?
↓
Re-verify when necessary
↓
Use in new research
This would make the memory layer especially useful for long-running research agents, where information changes over time.
I don't think the Perplexity API needs to become a memory database. An optional interface or integration point could allow developers to connect their own memory/experience backend.
I've been experimenting with this concept in CogniCore, an open-source cognitive infrastructure project for AI agents:
https://github.com/cognicore-dev/cognicore-my-openenv
CogniCore explores episodic, semantic and procedural memory, reflection, outcome-based learning, and experience transfer between agents.
I'd be interested in feedback on whether portable, provenance-aware research experience would be useful for Perplexity-powered applications.
The goal isn't just to remember what an agent found. It's to let the next agent start from what the previous agent already learned — while still verifying information that may have changed.
I'd like to propose an optional experience-memory layer for applications built with the Perplexity API.
The idea goes beyond storing conversation history. When an agent performs research, it often discovers useful information, failed approaches, source preferences, and conclusions that could be valuable to a future agent or research session.
Instead of making the next agent repeat the same research from scratch, validated experience could be transferred:
Research Agent A
↓
researches a topic
↓
sources + findings + failed approaches
↓
validated experience
↓
Memory
↓
Research Agent B
↓
related question
↓
retrieves relevant experience
↓
continues from previous work
For example, Agent A might discover that several sources are outdated, identify the most reliable sources, and verify a particular finding.
Instead of transferring the entire research conversation, the memory could preserve:
Research question
Relevant findings
Sources and provenance
Failed research approaches
Why a source/approach was rejected
Verification status
Timestamp
Relevant context
Confidence
I think failed research experience is particularly valuable.
Knowing "this source was outdated as of X" can prevent another agent from repeating the same mistake.
The receiving agent could first check whether the stored experience is still applicable before relying on it:
Stored experience
↓
Provenance check
↓
Freshness/context check
↓
Relevant?
↓
Re-verify when necessary
↓
Use in new research
This would make the memory layer especially useful for long-running research agents, where information changes over time.
I don't think the Perplexity API needs to become a memory database. An optional interface or integration point could allow developers to connect their own memory/experience backend.
I've been experimenting with this concept in CogniCore, an open-source cognitive infrastructure project for AI agents:
https://github.com/cognicore-dev/cognicore-my-openenv
CogniCore explores episodic, semantic and procedural memory, reflection, outcome-based learning, and experience transfer between agents.
I'd be interested in feedback on whether portable, provenance-aware research experience would be useful for Perplexity-powered applications.
The goal isn't just to remember what an agent found. It's to let the next agent start from what the previous agent already learned — while still verifying information that may have changed.