Lossless Working Memory Runtime

Context

The memory-md-temporal-head spec defines MEMORY.md as a rendered temporal head. This contract defines how that head behaves at runtime.

The goal is not a periodic report. The goal is a live working-memory system where one agent can move across many sessions, branches, and harnesses while keeping a single coherent head.

Core Runtime Invariant

One agent, many sessions, one head.

• different sessions and branches for the same agent write into the same temporal memory body
• MEMORY.md is the single rendered head over that body
• LCM patterns drive refinement over time, regardless of whether the source event was a prompt, a session end, or a compaction event

The rendered head must be both:

1. a dense operator-facing context buffer for immediate reinjection
2. a machine-facing index that points back to transcript, summary, and compaction lineage

Temporal Ownership and Scoping

All temporal state must remain scoped by agent_id and session_key where applicable.

• live transcripts are written with agent/session scoping
• session summaries inherit the same scoping
• compaction nodes inherit the same scoping
• lineage surfaced through MEMORY.md must preserve those identities
• identical session_key values across different agents must not collide, overwrite, or block transcript and summary persistence

Multiple instances of the same agent may write into the same temporal index. Runtime must therefore use merge-safe write protection, lease semantics, or equivalent conflict control so concurrent updates do not silently clobber the rendered head.

Live Session Behavior

Session start

At session start, context is injected immediately from:

• MEMORY.md
• other workspace identity files
• any harness-specific instruction files already supported

The point of MEMORY.md is continuity. Starting /new should still let the agent pick up where it left off.

Session in progress

While a conversation is active:

• transcript content is written to the database in real time
• transcript rows become queryable as soon as they are persisted
• transcript content may be embedded eagerly when runtime allows
• structural distillation continues separately through the existing entity/aspect/attribute pipeline

Transcript retrieval is not the primary retrieval path. It is a fallback source used when structured traversal cannot yet fire because the session has not been fully distilled, or when transcript-specific query behavior is explicitly needed.

Session End Behavior

When a session ends:

• a session summary is written and attached to transcript lineage
• the summary becomes a temporal artifact in the DAG
• MEMORY.md is updated with a shorter rendered reflection of the useful state change
• the rendered head keeps references back to the deeper temporal nodes

The result should be dense and rich enough that the next new session can resume with minimal context loss.

Compaction Behavior

If a session does not end cleanly and instead compacts:

• lossless compaction methodology produces a compaction artifact
• that artifact is stored as a first-class temporal node
• it feeds the same LCM/DAG refinement path as ordinary session-end summaries
• MEMORY.md remains the same head rather than forking into a separate compaction-only state

In other words, compaction is not a different memory system. It is another input path into the same one.

Retrieval and Distillation

Throughout the session lifecycle:

• entities, aspects, and attributes are distilled continuously
• duplicates and stale structural facts are maintained through the distillation pipeline
• prompt-submit retrieval should prefer the structured memory body when available
• transcripts are the fallback source until structural state catches up

This keeps prompt-time retrieval focused on the most relevant context without making raw transcripts the default source of truth.

Harness Realization

Harnesses will realize this model with different fidelity depending on:

• which hooks exist
• whether prompt-loop control is available
• whether compaction can be intercepted
• whether transcripts can be streamed or only flushed at checkpoints

The ideal model stays constant. Each harness maps onto it as closely as its runtime allows while preserving compatibility and scoping invariants.

Relationship to MEMORY.md

MEMORY.md is the universal context buffer for the agent, but it is not the only stored state.

• it always references deeper temporal artifacts
• temporal artifacts must remain discoverable from it
• its visible section stays concise enough for reinjection
• its index section preserves drill-down handles into the temporal body

That is the working model going forward.