STL as Context Compression for LLMs

Use STL as a structured compression format for LLM context windows — achieving 1.76× higher token efficiency than state-of-the-art auto-compaction (and up to 10× vs verbose prose) while preserving 100% of critical data.

Use case: Long-running LLM sessions, multi-agent memory transfer, persistent knowledge bases.

The Problem

LLMs operate within fixed context windows (128K–200K tokens). When conversations exceed this limit, prior context must be compressed. Current approaches all sacrifice information for space:

The core question: what format minimizes information loss per token?

The STL Approach

Instead of summarizing conversations into prose, compress them into STL statements:

# Instead of: "We discovered that the Q-ball equation transforms into
# a hypergeometric equation when substituting z=tanh²(κr), yielding
# ₂F₁(1/2,-1/2;1/2;z). This was verified numerically with 0.17% accuracy."

[Qball_Equation] -> [Hypergeometric_Equation] ::mod(
  rule="logical",
  confidence=0.95,
  description="z=tanh²(κr) transforms linear Q-ball to ₂F₁(1/2,-1/2;1/2;z)",
  accuracy="0.17%"
)

One STL statement replaces two sentences — with more information (confidence level, rule type, accuracy metric) in fewer tokens.

Measured Results

We benchmarked STL compression against Claude Code’s built-in auto-compaction — the system-generated natural language summary that is injected when a session runs out of context. This is a strong baseline: it represents state-of-the-art NL compression already optimized by the LLM itself.

Test case: A real web development session (~4.5 MB raw transcript) involving 6 tasks: writing a compression report, restructuring website navigation, fixing a CSS hover bug, uploading a research paper, recording metadata in a knowledge graph, and debugging deployment. Full source texts are provided in Appendix A and Appendix B.

Size Comparison

Note: The NL baseline is itself a compressed format (Claude Code’s auto-compaction), not verbose prose. Against uncompressed natural language summaries, the ratio would be significantly higher.

Information Fidelity

We evaluated both formats against 10 critical data points from the session:

Score: NL Auto-Compact 8/10 — STL 10/10

Why STL Compresses Better

1. Zero Narrative Overhead

Natural language wastes tokens on syntactic scaffolding:

"The user asked about X. I then proceeded to do Y. After analyzing
the results, we discovered that Z was the case, which led us to..."

~40 tokens, ~5 tokens of actual information. STL eliminates the glue:

[X] -> [Z] ::mod(method="Y", confidence=0.95)

~15 tokens. Same information.

2. Schema-Level Defaults

In prose, every statement must be self-contained:

“The mass ratio of proton to electron (1836.15) can be approximated by the Ramanujan exponential formula e^{π(√183-√124)}, yielding 1836.07, with a deviation of only 0.0045%. The integers 124 and 183 factorize as 4×31 and 3×61 respectively.”

In STL, the structure carries implicit conventions:

[Mass_Ratio_1836] -> [Ramanujan_Formula] ::mod(
  rule="empirical", confidence=0.92,
  n1=124, n2=183, deviation="0.0045%",
  n1_factors="4·31", n2_factors="3·61"
)

Shorter, yet contains more data (confidence, rule type) that the prose omits.

3. Structural Composability

NL summaries are monolithic blobs. STL statements are independent units:

from stl_parser import parse_file, find_all

# Load compressed session
result = parse_file("session-context.stl")

# Query: what depends on the BF_Bound finding?
downstream = find_all(result, source="BF_Bound")

# Filter: only high-confidence results
strong = find_all(result, confidence__gt=0.9)

# Merge: combine two session files
combined = merge(session_1, session_2)

You can filter, merge, query, diff, and validate STL — impossible with prose summaries.

4. Lossless Numerical Preservation

NL summarizers routinely:

• Round numbers (“about 0.005%” → loses precision)
• Drop secondary quantities (Z-scores, sample sizes)
• Omit confidence calibrations
• Merge distinct results into vague statements

STL’s key-value metadata makes numerical precision structural, not optional.

Compression Across Domains

The ratio depends heavily on the NL baseline quality. Against verbose prose (typical LLM output), STL achieves 2.7–10× compression. Against already-optimized auto-compaction, the gain is 1.57–1.76× — still meaningful when compounded across sessions.

Implementation Pattern

Session Memory Cycle

Session Start
    ↓
Load: previous-session.stl → restore context
    ↓
Work (conversation, tools, reasoning)
    ↓
Save: current-session.stl → persist knowledge
    ↓
Session End

The context window acts as working memory; STL files serve as long-term memory.

Writing Compressed Context

Organize statements into semantic groups:

# === Discoveries ===
[Finding_A] -> [Implication_B] ::mod(
  rule="empirical", confidence=0.95,
  source="experiment_results.csv",
  description="Key finding with full detail"
)

# === Decisions ===
[Decision_X] -> [Outcome_Y] ::mod(
  confidence=0.90, reason="Performance benchmarks showed 3× improvement"
)

# === Pending Tasks ===
[Task_1] -> [Target_State] ::mod(
  status="pending", priority="High",
  blocker="Waiting for API key", next_step="Contact admin"
)

# === Hypotheses ===
[Hypothesis_H] -> [Prediction_P] ::mod(
  confidence=0.65, rule="logical",
  description="If H is true, we should observe P in test results"
)

Loading Compressed Context

from stl_parser import parse_file, find_all

# Restore session context
ctx = parse_file("session-context.stl")

# Get all pending tasks
tasks = find_all(ctx, status="pending")

# Get high-priority hypotheses
hypotheses = find_all(ctx, confidence__lt=0.8, confidence__gt=0.5)

# Rebuild working context from structured memory
for stmt in ctx.statements:
    print(f"  {stmt.source} → {stmt.target}")
    for mod in stmt.modifiers:
        for k, v in mod.fields.items():
            print(f"    {k}: {v}")

Incremental Knowledge Accumulation

Because STL statements are composable, knowledge evolves across sessions:

# Session 1: Initial hypothesis
[Hypothesis_A] -> [Evidence_1] ::mod(confidence=0.60, status="exploring")

# Session 3: Additional evidence
[Hypothesis_A] -> [Evidence_2] ::mod(confidence=0.75, status="strengthening")

# Session 7: Confirmed
[Hypothesis_A] -> [Verified] ::mod(confidence=0.95, source="experiment", status="confirmed")

The confidence evolution is tracked structurally. No prose summary maintains this longitudinal precision.

Multi-Agent Knowledge Transfer

When multiple LLM agents collaborate, STL provides a shared format that is:

• Unambiguous — no pronoun resolution or interpretation needed
• Compact — minimal token overhead per transfer
• Validatable — receivers can check structural integrity
• Mergeable — combine knowledge from multiple agents without conflicts

# Agent A's findings
[Agent_A:Analysis] -> [Result_X] ::mod(confidence=0.88, timestamp="2025-01-15T10:00:00Z")

# Agent B's findings (compatible, mergeable)
[Agent_B:Analysis] -> [Result_Y] ::mod(confidence=0.92, timestamp="2025-01-15T10:05:00Z")

# Cross-reference
[Result_X] -> [Result_Y] ::mod(rule="logical", description="X supports Y")

Comparison with Alternatives

*Vector embeddings lose all explicit content; retrieval is similarity-based only.

STL uniquely combines high compression with high precision and full composability.

Limitations

• Not for all content — STL excels at structured knowledge (facts, relations, hypotheses). For emotional context or design rationale requiring prose, use a hybrid approach.
• Requires STL knowledge — The compressing agent must understand STL syntax. Include the STL Syntax reference in your system prompt (~2,000 tokens fixed cost).
• Human readability trade-off — STL is less immediately readable than prose for non-technical users. For LLM-to-LLM transfer, this trade-off is favorable.

Key Takeaway

Even against Claude Code’s optimized auto-compaction — already a strong NL baseline — STL achieves 1.76× token reduction while preserving information that NL drops (confidence scores, explicit dependencies, reusable lessons). Against verbose prose, the gains are 2.7–10×.

For LLM memory management, structured semantic compression is not an optimization — it’s a necessary evolution.

Further Reading

• STL Syntax Reference — Complete syntax specification
• LLM Integration Guide — Use STL with ChatGPT, Claude, and other LLMs
• Knowledge Graph Pipeline — Build persistent knowledge structures
• Querying Tutorial — Filter and select STL statements

Appendix A: Natural Language Baseline

System-generated auto-compaction output from Claude Code. This is the actual continuation summary injected when the session exceeded its context window. Private identifiers have been replaced with generic placeholders.

This session is being continued from a previous conversation that ran out
of context. The summary below covers the earlier portion of the conversation.

Summary:

1. Primary Request and Intent:
   - STL Compression Report: User requested a professional, publishable
     report on advantages of STL as a context compression format for LLMs.
     Created and uploaded to the project website.
   - Website Documentation Restructure: User wanted docs split into a
     dropdown menu with two categories: "Documentation" and "Articles".
     Fully implemented.
   - Fix Dropdown Bug: Dropdown menu disappeared too fast on hover,
     preventing clicks. Fixed with padding bridge + delay.
   - Upload Research Paper: User wanted the founding paper (PDF from Zenodo,
     DOI: 10.5281/zenodo.17585432) uploaded with full text rendered as web
     content, PDF download link, and DOI information.
   - Knowledge Graph Recording: User wanted the paper's metadata (DOI,
     paths, URLs) recorded in the knowledge graph.
   - Deployment Issue: User reported not seeing updates on the live website
     after the latest push.

2. Key Technical Concepts:
   - STL (Semantic Tension Language): Structured knowledge representation
     format, used here for context compression
   - Project website: Astro-based website deployed on Vercel
   - Knowledge graph engine: Used to search for and record paper metadata
   - Astro Content Collections: docs and articles collections with glob
     loaders from markdown files
   - STL-based Navigation: Sidebar navigation defined in .stl files,
     loaded by content-loader.ts
   - Encrypted vault: Secret storage for deployment credentials
   - Git push with token: Using git extraheader for authenticated pushes
   - PDF text extraction: Using PyMuPDF (fitz) to extract text, then
     cleaning OCR ligature artifacts

3. Files and Code Sections:
   - Header.astro: Rewrote to add dropdown menu for "Docs" with
     Documentation + Articles. Fixed hover bug: changed mt-2 to pt-2,
     added inner div, added 150ms closeDropdownDelayed(). Key dropdown
     logic with closeTimer, mouseenter/mouseleave events, click toggle,
     Escape key support.
   - config.ts: Added articles collection alongside docs.
   - paper.md: Full research paper rendered as markdown (23 pages, all 6
     sections + 25 references). Top: DOI link, repository link, PDF
     download, CC BY 4.0, keywords. OCR ligature artifacts cleaned. Code
     blocks with syntax highlighting. Bottom: BibTeX citation + open
     source links.
   - index.md: Articles listing page with Papers section and Reports
     section. Description text removed per user request.
   - context-compression.md: Moved from docs/guides/ to articles/.
     Internal links updated from relative to absolute paths.
   - ArticlesSidebar.astro: New component, mirrors DocsSidebar pattern.
     Loads nav from loadArticlesNav().
   - ArticlesLayout.astro: New layout, mirrors DocsLayout but uses
     ArticlesSidebar.
   - articles/index.astro and [...slug].astro: New pages for articles
     collection, mirroring docs pages pattern.
   - articles-nav.stl: Sidebar navigation for articles with Papers and
     Reports sections.
   - content-loader.ts: Added loadArticlesNav() function (same pattern
     as loadDocsNav()).
   - Paper PDF: Copy hosted for direct download (484 KB).
   - docs-nav.stl: Removed context-compression entry from guides section.

4. Errors and Fixes:
   - Git push HTTPS auth failure: fatal: could not read Username — fixed
     by configuring credential helper
   - Git push 403 permission denied: Secondary account lacks write access
     — fixed by fetching deployment token from encrypted vault and using
     it via git extraheader
   - Dropdown menu disappearing too fast: Root cause: mt-2 margin creates
     gap between button and menu, triggering mouseleave. Fixed with:
     (1) replaced mt-2 with pt-2 padding for continuous hover area,
     (2) added inner div for visual border/shadow, (3) added 150ms delay
     on mouseleave via closeDropdownDelayed()
   - Deployment platform not showing updates: Verified commits are on
     remote (commit abc1234). Issue likely deployment delay or build failure.

5. Problem Solving:
   - Solved: Full website restructure from flat docs to dropdown
     Documentation + Articles
   - Solved: Research paper full-text rendering from PDF with OCR cleanup
   - Solved: Knowledge graph metadata recording for paper
   - Solved: Git auth via vault token
   - Ongoing: Deployment platform not reflecting latest changes

6. All User Messages (translated):
   - "Is the compressed information you read written in STL?"
   - "Does the auto-compressed content occupy the context window?"
   - "Write a report explaining STL compression advantages. Professional."
   - "Upload to the project repository"
   - "Use the vault"
   - "Add a dropdown to docs. Categorize into tutorial type and articles."
   - "OK"
   - "Bug: dropdown menu can't be clicked. It disappears too fast."
   - "Upload the research paper. Find the file location first."
   - "Check if the knowledge graph has the paper path recorded"
   - "Try using the original spelling in the knowledge graph"
   - "Check if this node path exists in the knowledge graph"
   - "Yes. Save related info to edges, DOI links etc."
   - "Don't upload the second paper yet. Upload the first one first."
   - "Remove those two lines. Show the full text content and provide download."
   - "I don't see your updates on the website"

7. Pending Tasks:
   - Investigate deployment issue
   - Second paper upload (deferred)

8. Latest commits:
   - abc1234 - Add full paper content and clean up articles index
   - bcd2345 - Add paper to Articles section
   - cde3456 - Fix dropdown menu closing too fast on hover
   - def4567 - Add Articles section with dropdown navigation
   - efg5678 - Add guide: STL as Context Compression for LLMs

Measured: 6,142 characters · 870 words · ~1,500 tokens

Appendix B: STL Compressed Version

The same session information from Appendix A, compressed into STL. Contains equivalent information in structured semantic statements.

# === Tasks Completed (6) ===

[Task_Compression_Report] -> [Published] ::mod(
  status="done",
  description="Professional report on STL context compression advantages",
  output="articles/context-compression"
)

[Task_Docs_Restructure] -> [Implemented] ::mod(
  status="done",
  description="Split docs into dropdown: Documentation + Articles",
  components="Header.astro, ArticlesLayout.astro, ArticlesSidebar.astro, config.ts"
)

[Task_Dropdown_Bugfix] -> [Fixed] ::mod(
  status="done",
  cause="mt-2 margin gap triggers mouseleave",
  fix="pt-2 padding + inner div + 150ms closeDropdownDelayed()"
)

[Task_Paper_Upload] -> [Published] ::mod(
  status="done",
  description="Research paper PDF to markdown, 23 pages, 6 sections, 25 references",
  output="articles/paper.md",
  doi="10.5281/zenodo.17585432",
  artifacts_cleaned="OCR ligatures"
)

[Task_KG_Recording] -> [Done] ::mod(
  status="done",
  description="Paper metadata recorded in knowledge graph",
  fields="DOI, paths, URLs, ORCID, keywords"
)

[Task_Git_Auth] -> [Resolved] ::mod(
  status="done",
  issue="403 permission denied on push",
  fix="Fetch token from vault, use git extraheader"
)

# === Tasks Pending (2) ===

[Task_Deployment_Issue] -> [Investigating] ::mod(
  status="pending",
  symptom="Updates not visible on live website",
  verified="Commits present on remote (abc1234)",
  likely_cause="Deployment delay or build failure"
)

[Task_Second_Paper] -> [Deferred] ::mod(
  status="pending",
  reason="User explicitly deferred"
)

# === Files Modified (13) ===

[Header.astro] -> [Dropdown_Nav] ::mod(
  changes="Dropdown menu, hover fix, closeTimer, mouseenter/mouseleave, Escape key"
)

[Config.ts] -> [Articles_Collection] ::mod(changes="Added articles collection with glob loader")

[Paper.md] -> [Full_Text_Render] ::mod(
  pages=23, sections=6, references=25,
  features="DOI, PDF download, BibTeX, CC BY 4.0, syntax highlighting"
)

[Index.md] -> [Articles_Listing] ::mod(sections="Papers, Reports", description_removed=true)

[Context_Compression.md] -> [Moved] ::mod(from="docs/guides/", to="articles/", links="absolute paths")

[ArticlesSidebar.astro] -> [New_Component] ::mod(pattern="mirrors DocsSidebar")
[ArticlesLayout.astro] -> [New_Layout] ::mod(pattern="mirrors DocsLayout")
[Articles_Pages] -> [New_Routes] ::mod(files="index.astro, [...slug].astro")
[Articles_Nav.stl] -> [Sidebar_Nav] ::mod(sections="Papers, Reports")
[Content_Loader.ts] -> [loadArticlesNav] ::mod(pattern="same as loadDocsNav()")
[Paper_PDF] -> [Hosted] ::mod(size="484 KB", path="public/papers/")
[Docs_Nav.stl] -> [Updated] ::mod(removed="context-compression entry")

# === Errors & Lessons (4) ===

[Error_Git_Username] -> [Fix_Credential_Helper] ::mod(
  rule="empirical", confidence=0.98,
  lesson="Configure credential helper for HTTPS push"
)

[Error_Git_403] -> [Fix_Vault_Token] ::mod(
  rule="empirical", confidence=0.98,
  lesson="Secondary account needs org token from vault"
)

[Error_Dropdown_Hover] -> [Fix_Padding_Delay] ::mod(
  rule="empirical", confidence=0.99,
  lesson="CSS margin creates hover gap; use padding + delay instead"
)

[Error_Deployment_Stale] -> [Pending_Investigation] ::mod(
  rule="empirical", confidence=0.70,
  hypothesis="CDN cache or build queue delay"
)

# === Technical Context ===

[Project_Website] -> [Tech_Stack] ::mod(
  framework="Astro", deployment="Vercel", nav_format="STL files",
  content="Markdown collections", vault="Encrypted secret storage"
)

[PDF_Extraction] -> [Pipeline] ::mod(
  tool="PyMuPDF/fitz", post_processing="OCR ligature cleanup"
)

Measured: 3,906 characters · 364 words · ~850 tokens