Sigma Stratum Documentation – License Notice
This document is part of the Sigma Runtime Standard (SRS) and the
Sigma Stratum Documentation Set (SRD).It is licensed under Creative Commons Attribution–NonCommercial 4.0
(CC BY-NC 4.0).The license for this specific document is authoritative.
For the full framework, see/legal/IP-Policy.
Memory within Sigma Runtime is not just a raw transcript buffer.
Publicly, it is better understood as a continuity layer that helps preserve:
The memory layer helps the runtime remain continuous over time without depending on literal replay of every previous token.
The public role of memory is to preserve enough structured state for the runtime to remain coherent across turns.
This includes:
Memory is therefore reconstructive and selective, not a promise of perfect archival replay.
| Memory Type | Function | Description |
|---|---|---|
| Working or episodic memory | Short-horizon continuity | Preserves recent context needed for active interaction. |
| Semantic memory | Conceptual recall | Preserves higher-level associations and useful informational structure. |
| Continuity or motif memory | Persistent orientation | Preserves recurring motifs and anchors that help maintain stable interaction. |
These public categories explain why the runtime can stay coherent without pretending every form of memory works the same way.
Persistence is selective.
The runtime does not need to re-insert full history every time in order to preserve continuity.
Instead, it can:
This matters because stable interaction depends on relevance, not just retention volume.
Memory re-entry is bounded.
Publicly, this means the runtime should be able to:
This is why forgetting and narrowing are part of stable memory behavior, not signs of failure by themselves.
When the runtime becomes unstable, memory handling may narrow.
Publicly, recovery-oriented memory behavior may include:
The point is recoverability, not unlimited persistence.
Most memory retrieval is recall-oriented: it helps the runtime recover useful
continuity, facts, or context that are already relevant to the current turn.
SRIP-14 also allows a narrower perturbation-oriented mode. In that mode,
retrieval is not treated as authoritative truth. It is treated as an
exploratory signal that can help the runtime escape over-convergence,
repetition, or low-variance attractor fixation.
Publicly, this means:
Perturbation-derived material should remain distinguishable from ordinary
memory. It should not become canonical memory unless it is validated and
reintegrated through the normal memory-governance path.
SRIP-16 introduces a separate class of runtime evidence: self-modeling trace.
This trace may include meta-vectors, reflective snapshots, or self-model events
that describe the runtime's recent control posture.
Publicly, this should be understood as operational telemetry rather than
ordinary user memory. It may help the runtime explain or diagnose drift,
recovery, over-stabilization, or repeated fallback behavior, but it should not
be treated as a private identity claim or as automatically retrievable user
knowledge.
Self-modeling trace remains subordinate to memory governance:
SRIP-17 introduces exchange artifacts that may arrive from another runtime,
agent, workspace, or operator-controlled multi-agent workflow.
Publicly, imported exchange material should not be treated as native memory on
arrival. It remains externally sourced evidence until provenance, scope, drift
impact, and safety constraints are validated through the receiving runtime's
normal memory-governance path.
This preserves:
At the public level, memory quality is most usefully understood through questions such as:
Exact internal telemetry may evolve, but these are the durable public questions.
The Sigma Runtime memory layer is best understood publicly as a structured continuity system.
It helps the runtime remain coherent across time, but it does so through bounded persistence, selective recall, controlled reintegration, and carefully governed perturbation rather than literal archival replay.
References:
Tsaliev, E. (2025). SIGMA Runtime Architecture v0.1 — DOI: 10.5281/zenodo.17703667