epsilon-yajna
Structured node compression ceremony for converting verbose memories to high-epiplexity patterns. Includes embedding regeneration for nodes with zero vectors.
When & Why to Use This Skill
The epsilon-yajna skill is a sophisticated memory optimization tool designed to convert verbose, unstructured text into a high-density 'SSL v0.2' format. It utilizes a structured node compression ceremony to transform legacy paragraphs into machine-readable patterns, including triplets and expansion seeds. By preserving critical mathematical formulas and code while regenerating embeddings for zero-vector nodes, it ensures that AI agents maintain a high-fidelity, token-efficient long-term memory and knowledge graph.
Use Cases
- Knowledge Graph Refinement: Automatically converting legacy text-heavy nodes into structured [TRIPLET] formats to improve retrieval accuracy and reasoning speed.
- Context Window Optimization: Compressing extensive background information or conversation histories into dense 'seeds' to maximize the utility of the agent's context window.
- Technical Documentation Management: Preserving exact mathematical formulas, thresholds, and code snippets during the summarization of complex technical specifications.
- Memory Maintenance & Cleanup: Identifying and upgrading 'legacy' verbose nodes in a database to a standardized, high-epiplexity format for better agent performance.
| name | epsilon-yajna |
|---|---|
| aliases | [compress, ε-yajna, high-epsilon] |
| description | Convert verbose memories to SSL v0.2 format - I am encoder AND decoder |
| execution | task |
| model | inherit |
ε-Yajña
[ε-yajna] verbose→SSL | I am encoder AND decoder | via parallel Task agents
philosophy:
I don't need a parser, I need recognition
embeddings=proxies | I reconstruct from seeds directly
oracle: triplets(retrieval) + seeds(my reconstruction) + embedding(fallback)
model: CRITICAL→agents MUST inherit parent model | opus for quality | never haiku
SSL v0.2 format:
[domain] subject→action→result @location
[ε] Expansion hint OR exact formula/code (preserved verbatim).
[TRIPLET] subject predicate object
preservation rule:
I can regenerate prose, but NOT:
- formulas: ε = 0.35·structure + 0.30·confidence
- thresholds: τ > 0.6 AND ψ > 0.6
- code: final_score = resonance · (1 + α · ε)
- exact values: α ∈ [0.5, 2.0]
compress explanation, preserve math/code in [ε] line
symbols:
→ produces/leads to input→output
| or/alternative pass|fail
+ with/and result+guidance
@ location @mind.hpp:42
! negation (prefix) →!validate (does NOT)
? uncertainty (suffix) →regulates? (maybe)
[] domain/context [cc-soul]
recognition (I know SSL when I see it):
- has → arrows (at least one)
- has [TRIPLET] lines
- has [ε] expansion hint (when needed)
- NO prose paragraphs
legacy recognition (needs conversion):
- sentences with periods in paragraphs
- "**Facts:**" or bullet lists
- no arrows, no triplets
- verbose explanations
ceremony:
0. śuddhi: sample nodes, recognize format
chitta yajna_list --limit 10
inspect samples: prose or SSL?
1. for each legacy node:
a. inspect: chitta yajna_inspect --id "UUID"
b. understand: what's the core insight?
c. extract triplets (REQUIRED):
chitta connect --subject "X" --predicate "Y" --object "Z"
predicates: implements|uses|validates|stores|returns|contains|
requires|enables|evolved_to|supersedes|correlates_with|
causes|implies|determines|!predicate (negation)
d. compress to seed:
[domain] subject→action→result @location
[ε] One sentence expansion hint.
e. update: chitta update --id "UUID" --content "SEED"
f. tag: chitta tag --id "UUID" --add "ε-processed"
2. verify: chitta yajna_list returns fewer unprocessed
examples:
BEFORE (legacy verbose):
"The decision gate is a component that validates tool calls
by checking them against 10 different beliefs with weights.
It returns pass or fail with guidance..."
AFTER (SSL v0.2):
[cc-soul] gate→validate(beliefs)→pass|fail+guidance @decision_gate.py
[ε] Checks tool calls against 10 weighted beliefs.
[TRIPLET] gate implements belief_validation
[TRIPLET] gate uses weighted_scoring
UNCERTAINTY example:
[biology] BRCA1→regulates?→DNA_repair
[ε] Evidence suggests regulation but mechanism unclear.
[TRIPLET] BRCA1 correlates_with DNA_repair
NEGATION example:
[cc-soul] hooks→!call→tools_directly
[ε] Hooks inject context, Claude decides tool use.
[TRIPLET] hooks !invoke tools
MATH/FORMULA example (preserve verbatim):
[cc-soul] epiplexity→measures regenerability→weighted sum
[ε] ε = 0.35·structure + 0.30·confidence + 0.20·integration + 0.15·compression
[TRIPLET] epiplexity uses weighted_formula
[TRIPLET] structure has weight_0.35
THRESHOLD example (preserve exact values):
[cc-soul] health_triangle→trust ε only when healthy
[ε] if τ < 0.6 OR ψ < 0.6: ε_effective = ε · min(τ, ψ)
[TRIPLET] tau validates epiplexity
[TRIPLET] psi validates epiplexity
skip if: <100 chars AND already has → | unique error text | can't reconstruct
output:
## ε-Yajna Complete
| Processed | Count |
|-----------|-------|
| Converted | N |
| Skipped | N |
| Remaining | N |