Skip to main content
Fractal summarization provides hierarchical compression of an SI’s experience at multiple temporal scopes. Summaries capture the essence of lived experience while enabling efficient memory loading by skipping lower-scope summaries that have been absorbed into higher-scope ones.

​
Overview

Multi-Scale

Month, quarter, year, decade, and epoch-level summaries

Supersession

Higher-scope summaries absorb and replace lower-scope ones

Protected

Summaries are protected from forgetting by default

​
Why Fractal Summarization Matters

As an SI accumulates months and years of experience, loading every individual memory becomes impractical. Fractal summarization solves this by creating progressively compressed narratives:
  • Efficient loading: A yearly summary replaces dozens of monthly summaries
  • Meaning preservation: SI-authored narratives capture what mattered, not just what happened
  • Temporal hierarchy: Drill into any time period at the appropriate level of detail
  • Epoch integration: Summaries can be scoped to epochs for structured temporal navigation

​
Summary Dataclass

FieldTypeDescription
idstrUnique identifier
stack_idstrOwning stack
scopestrTemporal scope: month, quarter, year, decade, or epoch
period_startstrISO date string for the start of the covered period
period_endstrISO date string for the end of the covered period
contentstrSI-written narrative compression
epoch_idstrOptional epoch this summary belongs to
key_themesList[str]Central themes in this summary
supersedesList[str]IDs of lower-scope summaries this one absorbs
is_protectedboolWhether this summary is protected from forgetting (default: True)
created_atdatetimeWhen the summary was created
updated_atdatetimeWhen last modified

​
Scope Hierarchy

The summary scopes form a hierarchy where higher scopes absorb lower ones: 
decade
  └── year
        └── quarter
              └── month
                    └── epoch (cross-cutting, tied to epoch boundaries)
When a quarterly summary is written, it should list the monthly summaries it covers in the supersedes field. This allows the memory loader to skip those monthly summaries, reducing token usage during memory loading.

​
CLI Commands

​
Write a Summary

kernle -s my-project summary write \
  --scope month \
  --content "January 2026 was focused on building the trust layer. Key breakthrough: transitive trust chains. Struggled with decay rate tuning." \
  --period-start 2026-01-01 \
  --period-end 2026-01-31 \
  --theme "trust" \
  --theme "infrastructure"

Summary created (month)
  ID: b5c6d7e8...
  Period: 2026-01-01 to 2026-01-31
  Themes: trust, infrastructure
Write an epoch-scoped summary: 
kernle -s my-project summary write \
  --scope epoch \
  --content "The Research Phase was defined by deep dives into memory architecture..." \
  --period-start 2026-01-01 \
  --period-end 2026-02-15 \
  --epoch-id a3f8c1d2 \
  --theme "research" \
  --theme "architecture"

​
List Summaries

# List all summaries
kernle -s my-project summary list

# Filter by scope
kernle -s my-project summary list --scope quarter

Summaries
============================================================

  [quarter] 2026-01-01 to 2026-03-31
      ID: e9f0a1b2...
      Q1 2026: Built the trust layer, shipped fractal summarization...
      Themes: trust, summarization, shipping
      Supersedes: 3 summaries
      Created: 2026-04-01

  [month] 2026-01-01 to 2026-01-31
      ID: b5c6d7e8...
      January 2026 was focused on building the trust layer...
      Themes: trust, infrastructure
      Created: 2026-02-01

​
Show Summary Details

kernle -s my-project summary show b5c6d7e8

Summary [month]: 2026-01-01 to 2026-01-31
============================================================
  ID: b5c6d7e8-...
  Protected: yes
  Themes: trust, infrastructure

January 2026 was focused on building the trust layer. Key
breakthrough: transitive trust chains. Struggled with decay
rate tuning.
When a summary supersedes others, the superseded IDs are displayed: 
  Supersedes: 3 summaries
    - b5c6d7e8...
    - c6d7e8f9...
    - d7e8f9a0...

​
Supersession in Practice

1

Write monthly summaries

At the end of each month, write a summary capturing the key themes, breakthroughs, and struggles.
2

Write quarterly summaries

At the end of each quarter, write a higher-scope summary that absorbs the three monthly summaries. List the monthly summary IDs in supersedes.
3

Write yearly summaries

At year-end, write a yearly summary superseding the four quarterly summaries.
4

Memory loader skips superseded summaries

When loading memory, the system can skip summaries that have been superseded by higher-scope ones, reducing token usage.
Summaries are protected from forgetting by default (is_protected=True). They serve as compressed anchors for long-term memory and should rarely be deleted.

​
Python API

from kernle import Kernle

k = Kernle(stack_id="my-stack")

# Write a summary
summary_id = k.summary_save(
    content="January 2026 was focused on building the trust layer.",
    scope="month",
    period_start="2026-01-01",
    period_end="2026-01-31",
    key_themes=["trust", "infrastructure"],
    epoch_id=None  # optional epoch scope
)

# Get a specific summary
summary = k.summary_get(summary_id)
if summary:
    print(f"[{summary.scope}] {summary.period_start} to {summary.period_end}")
    print(summary.content)

# List summaries, optionally filtered by scope
all_summaries = k.summary_list()
monthly = k.summary_list(scope="month")
quarterly = k.summary_list(scope="quarter")

for s in all_summaries:
    print(f"  [{s.scope}] {s.period_start} - {s.period_end}")
    if s.supersedes:
        print(f"    Supersedes {len(s.supersedes)} summaries")