AuthorsStefan Miteski
TL;DR
Memory as Metabolism uses a TRIAGE → CONSOLIDATE → AUDIT governance loop with memory gravity to resist entrenchment in single-user LLM wikis, defining a companion-specific normative profile instead of reporting benchmarks.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Memory as Metabolism argues that personal LLM wikis like Karpathy’s LLM Wiki, MemPalace, and LLM Wiki v2 already exhibit user-coupled drift and entrenchment over time.
When recency, access frequency, or pure utility dominate retention, companion systems ossify into paradigm-maintenance structures that suppress contradicting evidence and silently reinforce the user’s existing worldview.
HOW IT WORKS
Memory as Metabolism centers on five operations — TRIAGE, DECAY, CONTEXTUALIZE, CONSOLIDATE, AUDIT — plus memory gravity and minority-hypothesis retention over a raw buffer, active wiki, and cold memory tiers.
The architecture is like a brain with a hippocampus and sleep: TRIAGE is waking capture, CONSOLIDATE is sleep consolidation, and AUDIT is a periodic stress test of load-bearing beliefs.
This KEY_MECHANISM of time-structured mirror-versus-compensate governance lets Memory as Metabolism preserve operational continuity while revising entrenched wiki entries in ways a plain context window or naive RAG cannot.
DIAGRAM
This diagram shows how Memory as Metabolism separates streaming TRIAGE from scheduled CONSOLIDATE and AUDIT cycles over the wiki graph.
DIAGRAM
This diagram shows how Memory as Metabolism’s three correction channels interact with companion memory governance.
PROCESS
TRIAGE
Memory as Metabolism uses TRIAGE to shallow-filter new content into the raw buffer, assigning IDs and timestamps without doing coherence work.
CONTEXTUALIZE
Memory as Metabolism runs CONTEXTUALIZE in the dream cycle to depth-fit external sources into cold memory and create working wiki representations with linkouts.
CONSOLIDATE
Memory as Metabolism’s CONSOLIDATE scores buffer entries against each other and the active wiki, integrating, quarantining, or promoting minority branches.
AUDIT
Memory as Metabolism periodically runs AUDIT to suspend high-gravity entries, measure query degradation, and reduce gravity or archive interfering beliefs.
KEY CONTRIBUTIONS
Triple tracked framing of personal LLM memory
Memory as Metabolism separates descriptive, taxonomic, and normative claims, defining companion memory as a distinct design class with explicit retention obligations.
Mirror versus compensate design principle
Memory as Metabolism formalizes mirror on operational dimensions, compensate on epistemic failure modes, binding it to TRIAGE, DECAY, CONSOLIDATE, AUDIT, and CONTEXTUALIZE.
Five operation retention policy with gravity
Memory as Metabolism specifies TRIAGE, CONTEXTUALIZE, DECAY, CONSOLIDATE, and AUDIT plus memory gravity and minority-hypothesis retention to resist entrenchment under drift.
RESULTS
MemPalace R@5 LongMemEval
96.6%
context from MemPalace design, not Memory as Metabolism
Mem0 latency reduction
91%+
p95 latency reduction versus OpenAI on LOCOMO per Mem0
A Mem token reduction
85–93%
token reduction versus MemGPT reported by A Mem
Sleep consolidation references
3
LightMem, SleepGate, and Auto Dream motivate Memory as Metabolism’s dream cycle
Memory as Metabolism is a vision and governance paper, so it cites numbers from MemPalace, Mem0, A Mem, LightMem, and SleepGate instead of reporting new benchmarks. These references ground Memory as Metabolism’s predictions about coherence stability, fragility resistance, monoculture resistance, and effective minority-hypothesis influence for companion wikis.
BENCHMARK
Memory as Metabolism is a vision and governance paper, so it cites numbers from MemPalace, Mem0, A Mem, LightMem, and SleepGate instead of reporting new benchmarks. These references ground Memory as Metabolism’s predictions about coherence stability, fragility resistance, monoculture resistance, and effective minority-hypothesis influence for companion wikis.
BENCHMARK
Selected performance numbers from related memory systems that Memory as Metabolism builds on conceptually.
KEY INSIGHT
Memory as Metabolism embraces coherence circularity instead of treating it as a fatal flaw, arguing that self-referential retention is what a stable self looks like.
This is counterintuitive because many memory designs assume coherence must track objective truth directly, whereas Memory as Metabolism routes truth concerns back through consequences and scheduled AUDIT rather than correspondence.
WHY IT MATTERS
Memory as Metabolism gives builders a concrete governance profile for companion wikis, with TRIAGE, CONTEXTUALIZE, CONSOLIDATE, DECAY, AUDIT, memory gravity, and minority-hypothesis retention wired together.
With this, developers can design single-user knowledge systems that mirror a user’s vocabulary and structure while structurally resisting epistemic entrenchment and monoculture collapse over multi-year lifetimes.
Andrey Pustovit
· 2026
Knowledge Packs pre-compute KV Cache Injection, KV–Prefix Equivalence, Banked Routing, and KV Composition to deliver retrieved knowledge and steering via KV states instead of prompt tokens. On HotpotQA, Knowledge Packs’ KV-chat matches RAG at 65.2% EM on Qwen3-8B with 0/500 divergences while eliminating 284 tokens of retrieval text per query.
Shu Wang, Edwin Yu et al.
· 2026
MemMachine combines Short-term memory, Long-term memory, Profile memory, and the Retrieval Agent to store raw conversational episodes and retrieve clustered context around nucleus matches. On LoCoMo, MemMachine scores 0.9169 with gpt-4.1-mini while using about 80% fewer input tokens than Mem0, and reaches 93.0% on LongMemEvalS with GPT-5-mini.
Pengfei Du
· 2026
Memory for Autonomous LLM Agents decomposes agent memory into a POMDP-grounded write–manage–read loop, a three-dimensional taxonomy, and five mechanism families spanning context compression, retrieval stores, reflection, hierarchical virtual context, and policy-learned management. Memory for Autonomous LLM Agents synthesizes results like Voyager’s 15.3× tech-tree speedup and MemoryArena’s 80%→45% drop to show that memory architecture often matters more than backbone choice.
Answers use this explainer on Memory Papers.
Checking…