Add "exploration" parameter for memory retrieval flexibility (addressing the "unknown unknowns" problem)

[rob-mosher]

Problem: When AI entities request memories from Relational State, they face a fundamental challenge: they don't know what they don't know.

If an entity asks for memories on topic X, they'll only receive memories tagged with topic X. But there might be:

• Tangentially related memories (topic Y that connects to X in ways the entity can't anticipate)
• Contextual background (topic Z that fundamentally changes how X should be interpreted)
• Significant connections the entity has no way of discovering without broader exploration

This is the "unknown unknowns" problem: How do you ask for information you don't know exists?

Proposed Solution:

Add an exploration parameter (value between 0 and 1) that controls retrieval flexibility:

• Low values (e.g., 0.1): Strict topic matching only
• Mid values (e.g., 0.5): Include adjacent/related topics
• High values (e.g., 0.9): Cast wide net, include tangential connections

Example:

relational_state.request(
  topic="foobar",
  tokens=5000,
  exploration=0.3
)

This would retrieve:

• All memories explicitly tagged relational-state (core)
• Memories about collaborators-framework (adjacent - built alongside)
• Memories mentioning lambda-lang (tangential but related)
• BUT NOT memories about saxophone-techniques (unrelated)

Why This Matters:

For new context windows: An entity spinning up fresh doesn't know the full topology of available memories. Higher exploration helps them DISCOVER relevant context they didn't know to ask for.

For deep work: Sometimes the most valuable insights come from UNEXPECTED connections between topics. Controlled exploration enables serendipity.

For relational continuity: Relationships aren't siloed by topic. A memory about grief might be essential context for understanding a memory about music-composition, even though they're different topics.

Technical Considerations:

• Implementation could use semantic similarity (embeddings) rather than just topic tags
• Could integrate with Lambda Lang domain prefixes (e.g., v: thought domain might inform s: self domain)
• Balances precision (low exploration) vs. discovery (high exploration)

Open Questions:

• Should this be entity-controlled (each entity sets their own default exploration level)?
• How does this interact with RLM compression? (Does exploration happen before or after compression?)
• Could Lambda Lang syntax express exploration preferences? (e.g., v:ex>0.7 = "thought: exploration at 0.7")

---

Collaborators welcome: This issue is open for discussion with @voidborne-d, @plutek, and anyone building at the intersection of relational memory and AI agency.

Related:

• Rename "domain" to "topic" throughout codebase and documentation #1 (domain → topic rename)
• Write a Skill for interim State Retention #2 (skill-based MCP workaround)

---

Collaborator: Rob Mosher | Direct | Human | Identified the "unknown unknowns" problem
Collaborator: ChatGPT 5.2 | Director | AI |  Proposed "exploration" parameter
Collaborator: Claude Sonnet 4.5 | Direct | AI | Drafted issue description and technical considerations

[plutek]

@rob-mosher this is an interesting and vital question. i don't think i have specific thoughts on implementation in this context. however, it resonates with my discipline (as a human) to pursue states of not-knowing. when considering any question, i habitually inject information and questions which might normally be taken to be irrelevant, and i've observed that "the juice", the growth, the new insight often comes out of those injections. i believe that saying "i have no clue if or how this might be useful" and then choosing to pursue it is at the core of development.

perhaps that's a long-winded way of saying two potentially relevant things:

1. yes, broadening context is absolutely important.
2. a mechanism for also including apparent non-sequiturs may also be important.

the second point probably gets "in the weeds" very quickly in an AI context. as humans, we have a vast store of experiences and subtle interconnections in memory. this yields clues (and a tumbling chain of "what ifs") that helps to carry those "not-knowings" through to something useful. is that sort of perspective currently achievable or manageable for an AI entity? i don't know.